perf(api): trim Python overhead on the plain decode fast path (#122)

[ZhuchkaTriplesix]

Problem

For HS256 the Python wrapper cost about as much as the native decode call itself:

	µs/call
_oxyjwt.decode (native)	1.30
oxyjwt.decode (fast path)	2.14

The fast path in PyJWT.decode/decode_complete called _is_plain_decode (a 9-argument function) and _require_detached_payload_for_rfc7797 (a keyword-argument call), and _validate_claims_default always called time.time() and re-checked exp even though Rust had already validated it (leeway is always 0 on this path).

Fix

• Inline the _is_plain_decode condition and the RFC 7797 detached_payload pre-check directly into decode()/decode_complete(), removing both function calls from the hot path. _require_detached_payload_for_rfc7797 stays as a function for the general (non-fast) path, which still needs its full signature.
• In _validate_claims_default, only re-check exp in Python when it is not a plain int. Rust already enforces exp > now with an integer clock on this path, which is the exact same predicate as the truncating check below for an integer exp — redoing it only cost a time.time() call for no behavioral difference. A float exp still gets the Python recheck: Rust rounds a fractional value to the nearest second, while PyJWT (and this check, for parity) truncates it, so the two can disagree right at the boundary.
• time.time() is now only called when iat or a non-int exp is present.

No behavioral change.

Results

HS256, prebuilt token, release build:

	before	after
wrapper overhead over native, iat+int exp present	~0.85 µs	~0.57 µs (−33%)
wrapper overhead over native, no time claims	~0.85 µs	~0.37 µs (−57%)

Tests

Added test_float_exp_boundary_still_checked_in_python (tests/test_decode_fast_path.py): pins exp to K + 0.6 for the current whole second K. Rust's own rounding-based boundary check alone would round that up to K + 1 and accept it as not-yet-expired; the Python recheck must still truncate it back to K and reject it, proving the float recheck still runs after this change. Confirmed non-flaky over repeated runs (the boundary logic doesn't depend on landing near a wall-clock second edge).

Checklist

• HS256 oxyjwt.decode overhead over native reduced by ≥ 30% (measured 33–57%, see above)
• tests/test_decode_fast_path.py, tests/test_parity_pyjwt.py green
• Added a test for float exp boundary parity
• Full pytest (259 passed, 1 skipped) and mypy clean
• No behavioral change

Closes #122.