execution/cache: avoid rebuilding cache views on direct reads

[yperbasis]

Follow-up to #23095. This is an optimization; it is not required for the correctness of that PR.

Problem

#23095 binds each cache view to the database state and immutable-files view pinned by its transaction. Most hot callers already construct a TemporalGetter once and reuse its bound views, but some direct read paths still repeat view construction:

• asOfStateReader calls SharedDomains.GetLatest for every commitment-domain read.
• SharedDomains.GetLatest and GetLatestContext derive both cache generations even when the requested domain can use only one cache.
• View construction inspects the transaction even when no relevant cache is attached.
• The single-domain calls to variadic TxNumsInFiles allocate when made through TemporalDebugTx.

The work is small, but it is paid on reads whose cache-hit path is otherwise about 50 ns.

Proposed change

• Bind one TemporalGetter when constructing an asOfStateReader. Reuse it when only txNum changes, and bind a new getter when a clone receives a different transaction.
• Return immediately when no relevant cache is attached.
• For direct reads, derive only the state-cache or branch-cache view needed by the requested domain. Code helpers need only the state-cache view.
• Consider a clearly named, non-variadic accessor for one pinned domain-files end, so generation construction does not allocate through the debug interface.
• Keep a benchmark that compares direct and bound cache hits.

Correctness constraints

• A view must always be derived from the transaction that performs the read.
• A state-cache generation must still contain the Accounts, Storage, and Code file ends, even when the current read targets only one of those domains. The cache has one shared generation.
• A branch-cache generation uses the Commitment file end.
• Transactions without exact domain frontiers must remain ineligible for the relevant cache.
• Worker clones must preserve context-based metrics, and clones using another transaction must not reuse the source transaction's view.
• Publication and token-revocation semantics must remain unchanged.

Prototype measurements

Indicative local results on darwin/arm64, Apple M2 Max:

Path	Current #23095	Prototype
Fresh getter construction	about 0.69 us, 96 B, 5 allocs	0.34-0.42 us, 80 B, 1 alloc
Direct cache hit	about 111 ns, 0 allocs	77-79 ns, 0 allocs
Bound-getter cache hit	about 49 ns, 0 allocs	about 49 ns, 0 allocs
No-cache unique miss	about 220 ns, 8 B, 4 allocs	about 116 ns, 0 allocs
Cache-backed unique miss	about 324 ns, 32 B, 5 allocs	about 295 ns, 24 B, 1 alloc

The main expected benefit is that commitment reads through asOfStateReader use the bound-getter path. Before merging, repeat the microbenchmarks on amd64 and measure a representative commitment workload; the extra interface method is worthwhile only if the allocation reduction remains useful outside the microbenchmark.