refactor: use common LDValueConverter and LDContextEncoder in AI SDK

[mattrmc1]

Summary

Switches the AI SDK (server-ai) to consume the shared LDValueConverter and LDContextEncoder from common (com.launchdarkly.sdk) and removes the SDK's internal copies. This is a behavior-preserving refactor — the interpolation and config-parsing output are unchanged; the logic now lives in one shared place.

⚠️ Merge order — do not merge before #187 is merged and released

This PR depends on the common utilities added in #187. Because common is a separately published artifact, #188 can only merge to main after:

1. feat: add LDValueConverter and LDContextEncoder to common #187 is merged, and
2. a common release containing LDValueConverter + LDContextEncoder is published, and
3. the AI SDK resolves that released common version.

Merging this before that release would leave server-ai referencing common symbols that aren't in a published artifact yet. Do not merge until the release dependency is satisfied.

Changes

• Interpolator — the private contextToMap / singleContextToMap methods are removed; the ldctx template variable is now built via LDContextEncoder.encode(context). The interpolation policy is unchanged (no HTML escaping, missing/null → empty string, ldctx merged last so it always wins).

// before
merged.put("ldctx", contextToMap(context));
// after
merged.put("ldctx", LDContextEncoder.encode(context));

• AIConfigParser — now imports and uses com.launchdarkly.sdk.LDValueConverter (for model.parameters, model.custom, and tool parameters / customParameters) instead of the SDK-local copy.
• Removed the AI SDK's internal LDValueConverter and its LDValueConverterTest; that coverage now lives in common.

Migration

None. All changes are internal to server-ai; the public API and the rendered output (config maps and ldctx template values) are unchanged.

Test plan

• server-ai module builds and tests pass against the released common version
• InterpolatorTest — existing cases still pass (interpolation parity preserved), including single- and multi-kind ldctx rendering
• AIConfigParser behavior unchanged for model.parameters / model.custom / tool params
• Shared converter/encoder coverage verified in common (LDValueConverterTest, LDContextEncoderTest)