refactor(field_formatter): use Spark transformer for field-name cache

[mike1o1]

Summary

Replaces the per-process-dictionary cache from #68 with a compile-time Spark transformer that persists every (field, formatter) pair via Spark.Dsl.Transformer.persist/3. Runtime lookup is one Spark.Dsl.Extension.get_persisted/2 call. Preserves the typescript_field_names/0 callback precedence over the DSL.

Changes

• New AshTypescript.Resource.Transformers.PersistFormattedFields — walks public attributes, relationships, calculations, and aggregates; persists (field, builtin_formatter) -> formatted_string at compile time. field_names DSL overrides are baked into the same map.
• New AshTypescript.Resource.Info.get_formatted_field/3 — exposes the persisted lookup.
• format_field_for_client/3 reads the persisted state on the hot path. Resources with typescript_field_names/0, non-builtin (MFA) formatters, and non-AshTypescript.Resource modules fall through to the existing runtime path unchanged.

Performance

Comparable to PR #68 with different tradeoffs across request shapes:

• Faster for typical Phoenix request shapes (single-record + small lists) — no per-request cache warmup. Single-record fetch in a fresh process: 1.91 μs vs PR perf(rpc): cut output formatting time on large list responses #68's 34.65 μs (~18× faster).
• ~10% slower for 1500-record bulk syncs where PR perf(rpc): cut output formatting time on large list responses #68's cache fully amortizes — mix run benchmarks/output_formatter.exs: 5.14 ms vs 4.65 ms.

Preserves the bulk of PR #68's improvement vs pre-#68 baseline. For a 1500-record response (mix run benchmarks/output_formatter.exs):

State	Time	Memory
Pre-#68 (no optimization)	71.87 ms	61.61 MB
PR #68 (process-dict cache)	4.65 ms	5.72 MB
This PR (Spark transformer)	5.14 ms	5.96 MB

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies