fix: keep S3 session parts ordered across writers

[cmun2]

nextPartName() separated same-millisecond appends by advancing the timestamp past the clock (Math.max(now, this.lastMs + 1)). Those borrowed milliseconds then belong to whoever writes next, so a second instance sharing the bucket stamps a part with a millisecond the first instance had already taken, and its part sorts inside the earlier burst.

const a = new S3SessionStore({ bucket, client })   // two processes mirroring
const b = new S3SessionStore({ bucket, client })   // the same session

// t = T
for (const type of ['a1','a2','a3','a4','a5']) await a.append(KEY, [{ type }])
// t = T+1
await b.append(KEY, [{ type: 'b1' }])
// t = T+2
await a.append(KEY, [{ type: 'a6' }])

await a.load(KEY)
// written: a1 a2 a3 a4 a5 b1 a6
// loaded:  a1 b1 a2 a3 a4 a5 a6

a's burst is named T … T+4, so b's part — stamped with the real T+1 — lands in position two. The rand suffix only breaks ties inside one millisecond; it cannot undo a name that claims a millisecond that has not arrived.

This matters because the transcript is what resume replays. Across 300 appends from two instances, 1747 of 44460 pairs written at different wall-clock times came back inverted:

instances	inversions before	after
1	0 / 44460	0 / 44460
2	1747 / 44460	0 / 44460
3	383 / 44460	0 / 44460

A single instance was always correct, which is why the conformance suite did not catch it — every case builds one store, and one writer is the case that works.

The fix keeps the timestamp honest and puts the burst counter in its own field:

part-{epochMs13}-{rand6}.jsonl  →  part-{epochMs13}-{seq6}-{rand6}.jsonl

The millisecond orders parts across writers, the sequence orders one instance's appends inside a millisecond, and the rand suffix breaks what is left. A clock that steps backwards keeps the last millisecond and advances the sequence, so an instance's own parts stay in write order — the property Math.max(now, lastMs + 1) was also providing.

listSessions() matches both name shapes, so a bucket written by an earlier version keeps its timestamps and still loads in order. Parts of the two shapes are only unordered relative to each other inside a single shared millisecond, where two writers have no ordering to preserve anyway.

Three tests added: a burst against a second writer on one bucket, a backwards clock, and listSessions() reading a part written before the sequence field. The first fails on main. The existing same-millisecond ordering test and the 13 conformance checks pass unchanged; 25 tests in s3/, and the Redis and Postgres adapters are untouched.