fix: keep lastLog cache in sync after append truncation failures

[anyasabo]

Problem

appendEntries can leave cached lastLog metadata stale when this sequence occurs:

1. follower truncates conflicting suffix,
2. writing replacement entries fails (StoreLogs error),
3. function returns without recomputing lastLog from durable store.

This creates cache/store divergence (r.getLastLog() vs LogStore).

How this can happen in concrete terms

A realistic failure chain in production:

• follower receives conflicting append from leader,
• DeleteRange succeeds (durable suffix removed),
• replacement append fails due to transient I/O error, ENOSPC, or backend log-store failure,
• node reports stale lastLog index/term from memory cache.

Existing mitigations and gaps

Mitigations that reduce probability:

• healthy disks and alerting on storage errors,
• robust log-store implementations.

Gap prior to this PR: even when storage returned an explicit append failure, cache was not reconciled.

Impact

Consistency of local bookkeeping degrades under storage fault:

• incorrect retry/backoff behavior potential,
• prolonged replication instability after fault,
• harder post-fault recovery due to stale in-memory metadata.

How we would notice in production

• Error log: failed to append to logs shortly after conflict truncation.
• Replication not converging despite apparent progress.
• Potential mismatch between on-disk last index and last_log_* stats/debug output.

Provenance / Preconditions

• Longstanding latent bug; code had explicit TODO noting stale lastLog risk in this path.
• TODO and path originate in early implementation (fdb8145f, 2016), with logging additions later.
• No special cluster configuration required; trigger requires truncation + StoreLogs failure.

What this PR changes

• Recomputes cached lastLog from durable store on StoreLogs failure.
• Adds regression test that reproduces truncation+append-failure path with injected log-store error.

Reviewer reproduction (live in-process Raft node)

Reproduce pre-fix behavior

1. Checkout parent commit:
   • git checkout 0d1b34c^
2. Bring this PR's regression test into that tree:
   • git checkout fix/append-truncation-store-failure-lastlog -- raft_test.go
3. Run:
   • go test -run "TestRaft_AppendEntriesStoreLogsFailureRefreshesLastLogAfterTruncate" -count=1 .
4. Expected pre-fix: test fails with stale cached lastLog (cache index > durable store index).

Verify fixed behavior

1. Checkout this branch (fix/append-truncation-store-failure-lastlog).
2. Run:
   • go test -run "TestRaft_AppendEntriesStoreLogsFailureRefreshesLastLogAfterTruncate" -count=1 .
3. Expected: pass; cached metadata matches durable store after failure.

Test plan

• go test -run "TestRaft_AppendEntriesStoreLogsFailureRefreshesLastLogAfterTruncate" -count=1 .
• go test -run "TestRaft_AppendEntry$|TestRaft_AppendEntriesStoreLogsFailureRefreshesLastLogAfterTruncate" -count=1 .