Compose MRTR input requests with task-backed tools

[KirschBluteX]

Summary

• share one internal input-required runner between Core MRTR backcompat and task-backed tool execution, so thrown and returned InputRequiredResult values follow the same retry limit, state replacement, and malformed-result rules
• route task-backed elicitation, sampling, and roots requests through the existing task outgoing-request interceptor and resume the tool after tasks/update
• cancel sibling input waiters when a request fails, observe late failures, and clean up subscribers on cancellation
• preserve native MRTR responses and legacy synchronous MRTR backcompat while covering state-only and multi-round retries
• document the implemented task/MRTR composition behavior

This implements Case 1 only. Case 2, mid-execution promotion through a DeferTaskCreation/PromoteToTaskAsync-style API, remains separate pending its public API and lifecycle design.

Partially addresses #1635

Testing

• dotnet build ModelContextProtocol.slnx --configuration Release --no-restore -m:1 --verbosity minimal (36 projects, 0 warnings, 0 errors)
• dotnet test tests/ModelContextProtocol.Tests/ModelContextProtocol.Tests.csproj --configuration Release --framework net10.0 --no-build --no-restore --filter "FullyQualifiedName~McpTaskStoreTests|FullyQualifiedName~MrtrServerBackcompatTests|FullyQualifiedName~MrtrInputRequiredExceptionTests|FullyQualifiedName~MrtrReturnedInputRequiredResultNativeTests|FullyQualifiedName~MrtrHandlerLifecycleTests" --verbosity minimal (77 passed)
• dotnet test tests/ModelContextProtocol.AspNetCore.Tests/ModelContextProtocol.AspNetCore.Tests.csproj --configuration Release --framework net10.0 --no-build --no-restore --filter "FullyQualifiedName~Mrtr_Backcompat_" --verbosity minimal (12 passed, 6 stateless variants skipped by design)
• dotnet format tests/ModelContextProtocol.Tests/ModelContextProtocol.Tests.csproj --verify-no-changes --no-restore --include <changed Core/Tasks test and source files> --verbosity minimal
• git diff --check

A broader Execution!=Manual test run was stopped after more than five minutes without a final result; no pass is claimed for that run.

The upstream workflows for this fork PR are currently action_required and need maintainer approval to run; this is not a CI failure.