fix session corruption

[fspmarshall]

Fixes an issue where session recording could become corrupted during multipart upload if two conditions were met simultaneously:

• A slice needed to be uploaded with padding applied (occurs when the event source is taking too long to push events).
• An error occurs on the first upload attempt.

In this case, the slice would have its reader() method called multiple times. The reader() method determines the length of the slice based on the length of the underlying buffer and appends padding as needed, noting the length of the payload and length of the padding in the slice header. On subsequent calls to reader() this logic would be re-run and would count the padding bytes as part of the payload, resulting in corruption of the header and preventing the session recording from being read. Attempts to read the session recording would fail with the error gzip: invalid header.

This PR updates the multipart upload logic to prevent this condition from being hit, and updates the session recording read logic to be able to read existing corrupted sessions. Test coverage has been added to catch future issues with paddings/retry conditions (neither of which were properly covered), and the test coverage has been added to the reader to ensure it can decode a fake corrupted session that was generated manually using some of our existing session recording test utilities (lib/events/testdata/corrupted-session).

This issue is loosely related to #45877 as the two issues tend to happen together since they are both triggered when the session recording system is under extremely high load.

changelog: fixed an issue where attempts to play/export certain session recordings would fail with gzip: invalid header.

[rosstimothy]

/excludeflake *

[public-teleport-github-review-bot]

@fspmarshall See the table below for backport results.

Branch	Result
branch/v14	Create PR
branch/v15	Create PR
branch/v16	Create PR