transport: canonicalize NaN floats on encode

[cargopete]

Closes #20.

The Component Model canonical ABI mandates a single canonical NaN representation for floating-point values (canonicalize_nan{32,64}). Previously f32/f64 were encoded with a raw byte copy (wasm-tokio's F32Codec/F64Codec), so a non-canonical NaN (signalling, or a payload-carrying NaN) would be transmitted verbatim rather than canonicalized.

This wraps the float codecs so that any NaN is canonicalized on encode:

• f32 → 0x7fc0_0000
• f64 → 0x7ff8_0000_0000_0000

Decoding is left lenient and accepts any NaN representation (as suggested in the issue), so the decode path simply delegates to the existing wasm-tokio codec.

Implementation

A small impl_canonical_nan_codec! macro generates a wrapper codec (CanonicalNanF32Codec / CanonicalNanF64Codec) that canonicalizes in its Encoder and delegates decoding to the inner wasm-tokio codec. The existing impl_copy_codec! registrations for f32/f64 are switched to the wrappers, so every encode path (scalar and list) funnels through the canonicalization.

Tests

Added canonical_nan_f32 / canonical_nan_f64 exercising the real Encode path:

• a signalling/non-canonical NaN encodes to the canonical bit pattern
• a negative NaN canonicalizes to the positive canonical NaN
• non-NaN values encode unchanged