fix(android-driver): bounded gRPC retry with channel rebuild on broken pipe

[proksh]

Why

After #3311 reverted the channel-level retries from #3290, transient socket drops between the worker and the Android emulator (UNAVAILABLE / Broken pipe in viewHierarchy / deviceInfo / etc.) surface as INFRA_ERROR and fail the job. #3290 hid these but caused 15-min hangs because gRPC's transparent retries shared the same dead channel and the 120s per-call deadline let a wedged RPC drag on.

This PR adds a bounded retry layer above the gRPC client and rebuilds the channel when the socket is dead.

What

• GrpcRetry.withRetry — a single retry helper used by AndroidDriver.runDeviceCall. Retries UNAVAILABLE only (3 attempts, 200 ms backoff). Hard ceiling of 30s total budget — no possibility of a 15-min hang. DEADLINE_EXCEEDED, INVALID_ARGUMENT, INTERNAL are never retried.
• Channel rebuild on broken pipe — when the failure carries an IOException cause (or "broken pipe" / "io exception" in the message), AndroidDriver swaps in a fresh ManagedChannel + stubs before the next retry. This is what enableRetry() couldn't do: it kept hammering the dead transport.
• contentDescriptor now routes through runDeviceCall — view-hierarchy was the call surfacing in alerts but had its own weaker retry path (1 retry, 1s sleep, no channel rebuild). The duplicate callViewHierarchy is removed; it now gets the same treatment as every other RPC.

Tests

10 new unit tests in GrpcRetryTest. Cover: happy path, retry on UNAVAILABLE, exhaustion, non-retryable codes (DEADLINE_EXCEEDED / INVALID_ARGUMENT / INTERNAL), broken-pipe (cause + message) → onBrokenPipe fires before retry, plain UNAVAILABLE does not rebuild, total budget cuts off retries even when attempts remain.

./gradlew :maestro-client:test — 121 passing, 0 failures.

Scope notes

The per-call deadline (withDeadlineAfter(120s) on blockingStubWithTimeout) is intentionally not changed in this PR — that's a broader behavioral change touching every RPC.