perf(pydantic_ai): simplify trace shaping

[Abhijeet Prasad (AbhiPrasad)]

Avoid pre-serializing Pydantic AI values in tracing helpers. Keep integration-side shaping focused on readable span payloads and materializing binary content into Braintrust attachments, while leaving general serialization to Braintrust logging.

Update private helper names and tests from serialize_* to shape_* to reflect the narrower responsibility.

I did a quick local microbenchmark against an approximation of the previous helper behavior using the pydantic-ai latest nox env.

 Results:

 ┌────────────────────────────────────┬──────────┬─────────┬────────────────────────┐
 │ case                               │ old      │ new     │ speedup                │
 ├────────────────────────────────────┼──────────┼─────────┼────────────────────────┤
 │ binary message shaping             │ 24.83 µs │ 5.45 µs │ 4.6x                   │
 ├────────────────────────────────────┼──────────┼─────────┼────────────────────────┤
 │ text messages x5                   │ 51.79 µs │ 7.31 µs │ 7.1x                   │
 ├────────────────────────────────────┼──────────┼─────────┼────────────────────────┤
 │ nested binary content part         │ 15.01 µs │ 4.86 µs │ 3.1x                   │
 ├────────────────────────────────────┼──────────┼─────────┼────────────────────────┤
 │ model response text                │ 13.84 µs │ 1.94 µs │ 7.1x                   │
 ├────────────────────────────────────┼──────────┼─────────┼────────────────────────┤
 │ model settings copy vs passthrough │ 0.79 µs  │ ~0 µs   │ effectively eliminated │
 └────────────────────────────────────┴──────────┴─────────┴────────────────────────┘