enh(Net,Foundation): multipart parsing performance and stream bulk-read optimization

[matejk]

Summary

• Fix OSS-Fuzz timeout in multipart parsing (Fuzzing Issue in MailMessage #5288) and add Content-Length bulk-read optimization (Poco Multipart parsing is 10x slower than its Boost/beat or restinio equivalent #4118)
• Add xsgetn() bulk-read override to BufferedStreamBuf and fix exception handling in UnbufferedStreamBuf::xsgetn()
• Bulk-read boundary scanning with PrependStreamBuf for the no-Content-Length path
• Modernize MailMessage, MultipartReader, and PartHandler with C++17 features

Measured Performance (200 x 500KB parts, ~97 MB)

Handler using StreamCopier::copyToString (production MultiPartHandler path):

Scenario	main	This PR	Speedup
With Content-Length	224 ms	9 ms	25x
Without Content-Length	227 ms	94 ms	2.4x

Handler using stream.ignore() (NullPartHandler):

Scenario	main	This PR	Speedup
With Content-Length	426 ms	211 ms	2x
Without Content-Length	424 ms	296 ms	1.4x

Small parts (60,000 parts, 5.5 MB):

Scenario	main	This PR	Speedup
NullPartHandler	31 ms	71 ms	0.4x (regression)

OSS-Fuzz #5288 test case (46 KB crafted message, 181 parts, 2086 Content-Type params):

main	This PR	Speedup
3,879 ms	7 ms (rejected: empty boundary)	554x

Small-parts regression explanation

The 60,000 small-parts regression (31ms -> 71ms) is from bulk sgetn over-reading past boundaries. Each boundary match requires storing over-read bytes in a holdback buffer and replaying them via PrependStreamBuf for the next part's header parsing. For tiny parts (~10 bytes content, ~100 bytes headers), the holdback management overhead exceeds the old line-by-line scanning cost. For realistic payloads (large file uploads), the bulk-read approach is significantly faster.

Side-effect performance improvements

BufferedStreamBuf::xsgetn (affects all BufferedStreamBuf-derived streams)

The xsgetn() override copies directly from the internal buffer instead of calling sbumpc() per byte. This benefits every consumer using istream::read() or StreamCopier on any BufferedStreamBuf-derived stream:

Stream class	Module	Expected benefit
InflatingStreamBuf	Foundation	High — gzip/zlib decompression via StreamCopier
DeflatingStreamBuf	Foundation	High — compression input path
CryptoStreamBuf	Crypto	High — bulk encrypt/decrypt
ZipStreamBuf + PartialStreamBuf	Zip	High — ZIP extraction
DigestBuf	Foundation	Medium — limited by hash CPU cost
MultipartStreamBuf	Net	Proven 25x with Content-Length

UnbufferedStreamBuf::xsgetn exception safety

Fixes StreamCopier::copyToString returning empty when reading through Base64Decoder on streams that throw mid-read (e.g., multipart without final boundary). Previously, an exception during xsgetn discarded all bytes read so far.

HTTP server impact

For HTTP multipart form uploads (HTMLForm::readMultipart):

Scenario	Impact
File uploads with Content-Length per part	25x faster content reading
File uploads without Content-Length (typical browser)	2.4x faster (bulk boundary scanning)
Small form fields	Negligible (< 1ms either way)
Plain HTTP body reading via StreamCopier	Moderate — xsgetn benefits any read()-based consumer on HTTPFixedLengthStream

Changes

Fix #5288: MailMessage multipart parsing timeout

• Cache multipart state in makeMultipart() — eliminates O(parts x content_type_params) re-parsing
• Validate empty boundary before entering multipart loop
• Add MailMessage::MAX_PARTS (100,000) limit as DoS protection
• Replace char-by-char stream drain with istr.ignore()

Fix #4118: Content-Length bulk-read optimization

• MultipartReader::nextPart() extracts Content-Length from part headers (with 32-bit overflow protection)
• readContent() uses sgetn() for bulk reading when Content-Length is known
• PatternMatcher validates boundary after Content-Length exhaustion (lazy-allocated only when needed)
• Increase STREAM_BUFFER_SIZE from 1 KB to 16 KB

Bulk-read boundary scanning (no Content-Length path)

• readContent() uses sgetn() to bulk-read into buffer, then scanForBoundary() searches in-buffer
• Over-read bytes stored in holdback vector, replayed via PrependStreamBuf (reusable, zero-copy string_view prefix)
• Persistent std::istream wrapper avoids per-part iostream construction overhead
• _readAhead as vector<char> with std::move between parts eliminates copies

BufferedStreamBuf::xsgetn() bulk-read override

• Copies directly from internal buffer via char_traits::copy + setg
• Delegates refilling to existing underflow() for consistent behavior

UnbufferedStreamBuf::xsgetn() exception safety

• Catches exceptions mid-read and returns bytes already copied
• Fixes StreamCopier::copyToString through Base64Decoder on truncated streams

C++17 modernization

• MailMessage: [[nodiscard]], = delete, structured bindings, std::none_of, if-with-initializer
• MultipartReader: = delete, explicit != nullptr, static constexpr, enum class
• PartHandler: = default constructor/destructor, = delete copy
• PatternMatcher, PrependStreamBuf: implementation-private classes in .cpp, string_view, [[nodiscard]]

Test plan

• All 23 MailMessage tests pass (7 new: empty boundary, Content-Length, zero-CL, 60K parts, 100K+ rejected, 200x500KB with/without CL)
• All 9 MultipartReader tests pass (including Unix line ends, no final boundary)
• All 15 HTMLForm tests pass
• All 102 Foundation Streams+FileStream tests pass
• Profiled with Apple Instruments — remaining hotspots are MessageHeader::read (3.1%), memory management (7.5%), our scanForBoundary+readContent (2.4%)
• Full CI