[fxfactorial]

Fixes #221.

Problem

Building against Go 1.27+ (or gotip after golang/go@388145c) breaks wsflate with:

wsflate: bad compressor: unexpected stream tail: 0x7d0a0300 vs 0x0000ffff

Root cause

Go 1.27 rewrote compress/flate (golang/go@388145c, golang/go#75532). The new encoder
ends a stream with an empty final fixed-Huffman block (BFINAL=1, BTYPE=01,
i.e. 03 00 00 — 3 bytes) instead of the empty non-final stored block
(00 00 00 FF FF — 4 bytes) that previous versions appended as a flush trailer.

cbuf holds back the last 4 bytes, so the held tail is now {0xff, 0xff, 0x03, 0x00}
and the old checkTail() rejected it. Verified byte-for-byte across all levels 0-9 and
various payloads: Go 1.26 -> tail 00 00 FF FF; Go 1.27 -> tail FF FF 03 00.

Both tails are valid per RFC 7692 section 7.2.3.4
("Using a DEFLATE Block with BFINAL Set to 1").

Fix

checkTail() in wsflate/writer.go now accepts both tails:

• Classic tail (00 00 FF FF): unchanged — the 4 held bytes are dropped (the
  receiver appends them back per RFC 7692 7.2.2).
• Go 1.27 tail (FF FF 03 00): the 4 held bytes are passed through to the
  destination so the on-wire payload is a complete, self-terminating DEFLATE stream
  (... [00 00 00 FF FF] [03 00]). The receiver then appends 00 00 FF FF per 7.2.2,
  which lands after the BFINAL block and is ignored by the decompressor.
• Any other tail is still rejected (error now lists both accepted tails).
• A tailChecked flag keeps checkTail idempotent across Flush() + Close()
  (as Helper.CompressTo does), and Reset() clears it.

The on-wire format for classic-tail compressors is byte-identical to before; only
Go 1.27's tail is newly accepted. The extra 1 byte of payload for the BFINAL tail is
the spec-sanctioned way to keep the stream byte-aligned and minimal.

Tests

• writer_tail_test.go: toolchain-independent unit tests — a blockCompressor
  drives the exact tail bytes, so the new BFINAL branch is covered even on Go 1.26.
  Covers both accepted tails, the rejected-tail case, and a default round-trip.
• zz_edge_test.go: round-trips for empty/short/exact-4/big payloads,
  fragmented writes with mid-message Flush() calls, and context-takeover
  (single long-lived writer, carried LZ77 window).

Verified on go1.26.8 and go1.27.0: full module test suite passes on both,
plus -race and go vet clean.

cc @gobwas @cristaloleg — thanks @prattmic for the analysis in #221 (this follows
the RFC 7.2.3.4 BFINAL approach discussed there).

[gobwas]

Hey @fxfactorial, thanks for your PR! I think for now we're considering #227, which feels to be more in line with the spec and less bound to a concrete compressor implementation behaviour (the stdlib's one): accepting a second known tail keeps us bound to how one particular compressor happens to encode its final block, which is exactly what broke here. #227 treats the message as finished at Flush() and keeps the compressor's Close output off, so no tail list is needed.