fix: preserve Anthropic streaming input and cache usage

[dgageot]

When Anthropic streams a response, input token counts and cache read/write
metrics arrive in the message_start event, while message_delta carries
only output tokens. The previous code built its usage snapshot from the
message_delta event alone, so input tokens and cache usage were silently
dropped from every streaming call.

Both the standard and beta adapter paths now map the SDK's accumulated
usage object (which the SDK merges across all events) instead of the
message_delta snapshot. Explicit delta counts, including zero, still
override the accumulated value, so the behavior for fields that arrive in
message_delta is unchanged. A raw-capture failure no longer suppresses
usage entirely — the code falls through to the accumulated snapshot.
track_usage opt-out and fresh per-call snapshots are maintained.

Tests cover: field omission from message_delta (input/cache preserved),
cumulative updates, zero-value replacement, streams with no message_start,
disabled tracking, and raw-capture failures.

Full task test passes with the two TestLoadExamples subtests for
dmr.yaml and unload_on_switch.yaml excluded — both require a local
Docker Model Runner that is not available in CI and fail identically on
the unmodified baseline.