Clean up profile and replay artifacts on pre-envelope drops

[adinauer]

Problem

SentryClient currently relies on lazy envelope-item serialization to clean up file-backed transaction profiles and replay recordings. If capture terminates after those sidecars exist but before envelope construction, serialization never takes ownership and the files are not promptly deleted.

This affects both EventProcessor and beforeSend* drop paths. PR #6142 adds new fail-closed exits when an EventProcessor throws, but the same ownership gap already exists for intentional processor drops and beforeSend* drops.

Affected paths

Profiled transactions

A profiled transaction can return before buildEnvelope(..., profilingTraceData) when:

• an EventProcessor returns null;
• an EventProcessor throws and the SDK fails closed;
• beforeSendTransaction returns null; or
• beforeSendTransaction throws.

The profile trace is normally deleted by the lazy serializer created in SentryEnvelopeItem.fromProfilingTrace. These early returns prevent that cleanup, allowing sensitive profile data and disk usage to accumulate in long-running Android apps or JVM services.

Session replays

A replay can return before replay envelope construction when:

• an EventProcessor returns null;
• an EventProcessor throws and the SDK fails closed;
• beforeSendReplay returns null; or
• beforeSendReplay throws.

Replay cleanup currently occurs in the serializer created by SentryEnvelopeItem.fromReplay. Bypassing it can retain the replay video or, for a backfilled replay, the entire replay folder.

Expected behavior

Introduce explicit ownership for file-backed sidecars so every terminal pre-envelope drop either:

1. transfers the sidecar to an envelope item that owns serialization and cleanup; or
2. disposes of the sidecar immediately at the capture boundary.

The implementation should:

• delete abandoned profile trace files;
• delete normal replay videos and the correct backfilled replay folder;
• record one dropped profile with the same discard reason as its transaction (event_processor, before_send, or callback_error);
• avoid double deletion and double accounting; and
• preserve existing replay-folder semantics and transaction/span discard quantities.

Tests

Add real temporary-file regression coverage for:

• profiled transactions dropped by processor null and processor exception;
• profiled transactions dropped by beforeSendTransaction null and exception;
• normal replay videos dropped by processor and beforeSendReplay paths;
• backfilled replay folders dropped by processor and beforeSendReplay paths; and
• DataCategory.Profile accounting under each corresponding discard reason.

Context

This was identified while reviewing #6142. The fail-closed processor behavior is correct; the issue is that cleanup is coupled to envelope serialization even though several valid terminal paths return before an envelope is built.

[linear-code]

JAVA-743

[adinauer]

Couple quick investigations around this:
Profiles should be cleaned up on next SDK init.
For profiles there doesn't seem to be a limit of how many can remain on disk and each can be up to 3MB (but can also just be a couple KB for short transactions).

Replay should be deleted when recording stops (after 1h) or on next restart (if never stopped).
From what I understand about replay, we have one video file per 5s (30-60 KB), up to a full hour (20-50MB).
Typically those files are deleted during envelope item serialization.