Preserve decimal precision in MCP tool arguments

[sylvesterkaczmarek]

Fixes #419.

Enable USE_BIG_DECIMAL_FOR_FLOATS on the private mapper used by BetaMcp. Decimal and exponent-form arguments then reach the MCP client without rounding, overflow to infinity, or underflow to zero.

Integral parsing remains unchanged. Decimal values inside the outgoing CallToolRequest map are now BigDecimal rather than Double; they continue to serialize as JSON numbers. Public method signatures and result conversion are unchanged.

Nine regression cases check precise decimals, values beyond Double's exponent range, nested lists/objects, a large integer control, strings/booleans/null/empty objects, numeric serialization and malformed-input rejection before invoking the client.

Validation

• New regressions on unchanged main: 6 failed, 3 passed.
• Complete anthropic-java-mcp test module: 36 passed, none skipped.
• MCP module formatKotlin, detektMain, detektTest and lintKotlin tasks passed.
• git diff --check passed.

./gradlew :anthropic-java-mcp:test \
  :anthropic-java-mcp:detektMain :anthropic-java-mcp:detektTest \
  :anthropic-java-mcp:lintKotlin --max-workers=1 --no-parallel

Tested on macOS arm64 with JDK 21 and the repository's locked dependencies. The MCP client is mocked at its request boundary; live MCP transport and the full multi-module SDK suite were not run. Existing Gradle/JVM warnings remain. No dependency or workflow changes.

[sylvesterkaczmarek]

The complete MCP module tests pass locally (36 tests), along with its formatting and detekt checks. Could a maintainer approve the pending CI workflow?

[tomer-ant]

Closing along with #419. The reasoning is in #419 (comment).