feat(async): chunk-level traversal (achunk-next) + AsyncReverseSeq termination-contract fix

[whilo]

First PR of the datahike cljs-async arc (PR A of the grounded workplan).

Chunk-level async traversal

PAsyncChunkedSeq/achunk-next on AsyncSeq: resolves [keys-array start end next] — the consumer runs a synchronous inner loop over the leaf's key array and awaits once per leaf, not per element. Element-wise anext allocates a continuation per element (measured ~4x slower on warm data in the datahike engine experiments); the chunk shape is what a query engine's scan kernels consume, and awaiting per 512-element chunk measured within ~8% of a fully synchronous loop.

Boundary semantics mirror the sync Iter's IChunkedSeq exactly: same-leaf right bound via path-get, next leaf via -next-path with the path-lt guard. Parity is pinned chunk-for-chunk against Iter (same boundaries, same elements) over full/mid/same-leaf/single/empty ranges plus a leaf-boundary sweep — both in-memory and through a restored async storage (cold node loads).

AsyncReverseSeq termination-contract fix

anext's body was a literal two-slot vector of when forms, so an exhausted reverse seq resolved a truthy [nil nil] instead of nil — one spurious nil element for any consumer written against AsyncSeq's contract (aseq/transduce's if-let, chunk walkers). All internal constructors happen to guard emptiness, so this was latent — but external consumers constructing or re-driving reverse seqs would hit it. The guard now encloses the whole tuple; the AsyncRange test helper had the same shape and is fixed alike; a regression test pins anext → nil on exhausted seqs of both orientations.

Also: the lookup facade gains the opts arity the other facades already have (datahike's IIndex/-lookup opts threading needs it).

Tests

node-tests (advanced): 137 tests / 1826 assertions green — includes the two new namespaces (async-exhaustion, async-chunk). cljs-only change; JVM untouched.