websocket: fix PMD end-of-message tail partial consumption (permessage-deflate)

[shinny-yangyang]

Root cause

• When finishing a permessage-deflate message, Beast always appended the 4‑byte RFC7692 tail (00 00 FF FF) but did not track how many of those bytes zlib had consumed. With very small avail_out (e.g. 1 byte), zlib could consume only part of the tail and the next iteration would re-append from the beginning, corrupting the stored block length and producing "invalid stored block length".

Fix

• Track per-message tail consumption (0..4) and only feed the remaining portion of the tail. After all 4 bytes are consumed, continue calling inflate with avail_in=0 to naturally finish. Reset the counter on new messages and at context takeover boundaries. Also treat zlib::error::need_buffers as non-fatal in the inflate wrapper.

Changes

• impl_base.hpp (deflateSupported=true): add pmd_type::rd_tail_in, reset in rd_deflated() and do_context_takeover_read(); add rd_tail_in()/rd_tail_in_add() helpers; clear need_buffers in inflate().
• impl_base.hpp (deflateSupported=false): add no-op stubs for rd_tail_in helpers.
• read.hpp: in both async and sync read paths, append only the unconsumed tail bytes, accumulate consumed bytes from zs.total_in, and finish the message when fin && zs.total_out == 0.
• example: add repro CMake target (example/repro), with server_pmd_writer and client_smallbuf to reproduce and validate the fix.

Validation

• Repro shows stable reception with 1‑byte reads and no inflate errors; 200 messages received cleanly and stream closes gracefully.

Compatibility

• Only affects PMD-enabled compressed reads; non-PMD and write paths unchanged.