perf(schema): index name-mapping lookups

[fallintoplace]

What changed

• Build one name lookup index for each ApplyNameMapping call.
• Reuse it for regular fields and list/map element, key, and value lookups.
• Keep aliases scoped to their parent mapping field, so nested fields with the same name still resolve correctly.
• Keep the exported NameMappingAccessor as the original empty, comparable type. The index is held by a private accessor used only by ApplyNameMapping.
• Add alias and API-compatibility regression tests.
• Add a wide-schema benchmark.

Why

Name mapping lookup previously scanned sibling fields for every schema field. On wide schemas, that repeated the same work and caused many allocations.

The new per-operation index builds the lookup once, then uses map lookups during traversal. It is private to each call and read-only after construction.

Benchmark

BenchmarkApplyNameMappingWideSchema with a 256-field schema on an Apple M1 Pro:

Metric	Before	After	Result
Time/op	0.90 ms	37 us	~24x faster
Memory/op	2.16 MB	68 KB	~97% less
Allocs/op	32,921	31	~99.9% fewer

Validation

• go test ./...
• go vet .
• Wide-schema benchmark with -benchmem