[9.3.0] Re-sample the execution graph clock converter per command

[fmeum]

Description

ExecutionGraphModule built its NanosToMillisSinceEpochConverter in a field initializer:

private NanosToMillisSinceEpochConverter nanosToMillis =
    BlazeClock.createNanosToMillisSinceEpochConverter();

Modules are instantiated once per server, so that converter captures the offset between the monotonic and the wall clock as it stood at server startup, and is then reused for every command that server ever serves. Build it in beforeCommand instead, so each command converts through an offset sampled during that command.

Motivation

System.nanoTime() and System.currentTimeMillis() are backed by different operating system clocks, and the offset between them is not constant. NTP steps the wall clock when it corrects it. A virtual machine does so more dramatically: shortly after boot, and again whenever a snapshot is resumed — while the monotonic clock never steps and does not advance across the pause.

A long-lived server that survives a snapshot resume keeps converting with an offset captured before the snapshot, so every start_timestamp_millis it writes into the execution graph log afterwards is shifted by the entire time the snapshot spent on disk. Anything that reconstructs a timeline from the log — critical path, action concurrency, correlation against other traces on the machine — is wrong by that amount. On bare metal an NTP step produces the same corruption at a smaller magnitude.

Build API Changes

No

Checklist

• I have added tests for the new use cases (if any).
• I have updated the documentation (if applicable).

Release Notes

RELNOTES: None

Closes #30594.

PiperOrigin-RevId: 960312424
Change-Id: I3295137f08c41bc6d76593766e4a8386dd548fb5

(cherry picked from commit 13e656c)

9.3.0 adaptation: the release branch has not had master's test refactor that merged createModule()/initializeRoots() into a single setUp() and dropped the UUID argument from startLogging, so the module.resetNanosToMillis() call was added to createModule() (where module is guaranteed to be constructed) rather than to setUp(), and the existing startLogging(eventBus, UUID.randomUUID(), ...) call sites were left as they are. The production change is unmodified.

Closes #30601