fix(ttstream): ignore late header frame after stream close instead of tearing down mux transport

[linhongyu510]

What type of PR is this?

• Bug fix

What this PR does / why we need it:

After a client stream is close()d, an in-flight header frame can still be
dispatched to onReadHeaderFrame. This is a benign late frame, but today it
surfaces as a fatal errUnexpectedHeader (code 12002), which the read loop
treats as a protocol error and uses to tear down the whole mux transport
connection (deferred Close(err)) — taking down every other stream sharing
the connection.

Root cause

close() parks a terminal signal (streamSigNone / streamSigInactive) into
the buffered headerSig channel (size 1) before marking the stream
inactive. A late header then:

1. sets s.header = fr.header,
2. tries to send streamSigActive into the now-full headerSig channel,
3. hits the select default branch, which returned errUnexpectedHeader.

Fix

In the default branch, when the stream is already streamStateInactive, the
header is an in-flight frame for an already-closed stream: drop it (s.header = nil) and return nil so the mux transport survives.

A genuine duplicate header (two header frames, no close in between) still
errors: it either hits the s.header != nil guard, or reaches the default
branch on an active stream.

Which issue(s) this PR fixes:

N/A (found via race-oracle testing; no linked issue).

Special notes for your reviewer:

• The added check is an atomic.LoadInt32(&s.state) == streamStateInactive,
  matching the existing state-machine convention in this file.
• Regression tests added in stream_client_lateheader_test.go:
  • Test_onReadHeaderFrameAfterCloseIsIgnored — late header after close is
    ignored, s.header not leaked.
  • Test_onReadDuplicateHeaderStillErrors — true duplicate still returns
    errUnexpectedHeader (guards against over-loosening).
  • Test_onReadHeaderFrameRaceWithClose — races close() against
    onReadHeaderFrame for 200 iterations; never produces a fatal error.
• Verified: on unmodified main, the late-header and race tests fail with
  12002 and the duplicate test still passes; with the fix all pass under
  -race. go vet clean; cross-compiles for linux/amd64, linux/arm64,
  darwin/{arm64,amd64}, windows/amd64.