fix: Make batch_add_requests split batches by serialized payload size (#953)

The 9 MB payload guard in `batch_add_requests` was inert:
`constrained_batches` was called without `get_len`, so the default
`len()` measured each request dict's key count (~4) instead of its
serialized size. Batches were therefore split only by the 25-request
count limit, and large requests shipped as one oversized POST that the
API rejects with 413, failing the whole call.

The guard now measures each request as its UTF-8 JSON byte length, using
the same serialization flags as the HTTP client's request body path. It
also passes `strict=False`, which preserves the previous contract for an
individually oversized request: it's sent in its own batch and left for
the API to judge, instead of raising a client-side `ValueError`. Both
the size-based splitting and the oversized-singleton path are covered by
new sync/async regression tests.

*✍️ Drafted by Claude Code*