feat: add ADR-conformant numeric coercion scenarios under @numeric-coercion-v2

Adds ADR-conformant numeric coercion coverage for the numeric coercion ADR, as a new opt-in suite so providers can adopt it as they migrate (following the opt-in @fractional-v1/@fractional-v2 precedent).

There is currently no numeric-coercion feature in the testbed, so this is purely additive; there is no prior suite to preserve.

• add a new opt-in tag @numeric-coercion with scenarios matching the ADR:
  • 10.0 fetched as Integer resolves to 10 (lossless coercion)
  • 3.14 fetched as Integer returns TYPE_MISMATCH
  • int variant within safe range fetched as Float widens losslessly
  • wide/boundary values (e.g. 9007199254740991) resolved via the canonical Long accessor (@long)
  • int variant exceeding int32 fetched as Integer returns TYPE_MISMATCH, tagged @int32-bounded (only for languages whose Integer accessor is int32; Go/Python and other int64/wide-Integer languages exclude it)
  • int variant exceeding int32 fetched as Float succeeds
• add the scenarios to both suites: the provider gherkin/ suite (real flagd container / wire) and the self-contained evaluator/ suite
• add a flags/numeric-coercion-flags.json (auto-merged into allFlags.json) plus the matching evaluator/flags fixtures
• register @numeric-coercion, @int32-bounded, @long, and the Long flag-value type token in agents.md and README.md
• bump FLAGD_BASE_IMAGE to the flagd version that ships the contract
• cut a new testbed release

Out of scope: load-time rejection of out-of-range values. The safe-integer cap is an ordinary schema validation (a warning in flagd today), not a hard load failure, so there is no "rejected at load" scenario here. The 2^53 - 1 boundary is covered by resolving it correctly via Long; hard-fail-on-out-of-range is a separate future feature.

Parent: https://github.com/open-feature/flagd/issues/1995

Start by comparing the existing opt-in @fractional-v1/@fractional-v2 scenarios in gherkin/ and evaluator/. Add the numeric-coercion scenarios and fixtures in flags/numeric-coercion-flags.json and evaluator/flags, then update agents.md, README.md, FLAGD_BASE_IMAGE, and the release metadata mentioned in the issue. Done means both suites cover the ADR cases and the new tags and Long token are registered.