fix: drop oversize AI events locally on the immediate path and align retry defaults

[eli-r-ph]

💡 Motivation and Context

Two v1 behavior fixes so the Rust SDK matches the other server SDKs.

The immediate AI variants now apply the 8 MiB per-event limit locally. The background AI lane already drops an event whose serialized properties exceed AI_MAX_EVENT_BYTES (strictly greater). capture_ai_immediate and capture_ai_batch_immediate sent such an event and relied on the backend to refuse it. Now prepare_immediate measures each event against a per-lane max_event_bytes. An event over the limit is not sent. It gets a local EventResult { result: Drop, details: "ai_event_too_big" }, which counts in submitted() and not_persisted() exactly like the backend's verdict would. If every event in a call is over the limit, no request is made. The analytics lane has no ceiling and its behavior is unchanged.

Retry defaults.

• max_capture_attempts now defaults to 4 (was 3).
• retry_initial_backoff_ms now defaults to 100 (was 200).
• Remote /flags retries no longer read the capture retry options. They use a fixed backoff of 300 ms, doubling up to 30 s, which matches posthog-python and posthog-go. feature_flags_request_max_retries still controls how many /flags retries happen.
• The exponential math is shared by one helper, exponential_backoff.

The migration guide's Retry section and a capture-retry-defaults changeset describe the new defaults. The guide and changeset text for the AI immediate drop lives in #280, which already rewrites those sentences. I pushed a follow-up commit there so the two PRs agree.

💚 How did you test it?

New and changed tests:

• prepare_immediate_drops_events_over_the_lane_ceiling_locally covers an event exactly at the ceiling (sent), one byte over (local drop), and every event over the ceiling (no request). It runs on both lanes.
• capture_ai_batch_immediate_drops_an_oversize_event_locally replaces the old "server returns too-big verdict" test, async and blocking. The mock only answers a request that carries the normal event and not the oversize one. The test asserts submitted() == 2, not_persisted() == 1, the local ai_event_too_big drop, and the normal event's Ok.
• feature_flags_backoff_doubles_from_300ms_up_to_30s covers the /flags schedule and its cap.
• The two existing /flags decision tests now assert the exact 300 ms wait while the capture backoff options are set to 1/5 ms, which proves /flags ignores them.
• backoff_exponential_growth_with_default_options now uses the builder defaults, so it pins 4 attempts and 100/200/400 ms.

Mutation checks. Each mutant fails at least one test, and each was reverted afterwards:

• AI lane max_event_bytes set to None: the async and the blocking integration tests fail.
• > changed to >=: the unit test fails.
• Defaults reverted to 3 attempts or to 200 ms: the default-options test fails.
• /flags pointed back at the capture backoff: both /flags decision tests fail.

Ran locally and passing:

• cargo fmt -- --check
• cargo clippy -- -D warnings
• cargo test --workspace
• cargo test --no-default-features
• cargo test --no-default-features --features error-tracking
• both test_tls_no_provider variants
• cargo test --features e2e-test,tls --no-default-features
• scripts/check-public-api.sh (no public API change)
• cargo package --locked

Trial merges into #279, #280 (including its new commit) and #285 are all clean. The combined #285 + #280 + this branch passes clippy and cargo test --workspace.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

🤖 Agent context

Written with Claude Opus 5.5 in Cursor, directed and reviewed by @eli-r-ph.