[Feature] MCP and HTTP tools, agentic tutorial, see-also pointers

[vmoens]

Stack from ghstack (oldest at bottom):

• -> [Feature] MCP and HTTP tools, agentic tutorial, see-also pointers #3737
• [Feature] ToolCompose with parallel dispatch, builtin tools, legacy adapter #3736
• [Feature] Agentic toolkit foundation: protocols, parsers, sandbox, REPL #3735

Adds the remaining built-ins, a sphinx-gallery tutorial, and soft
"see also" pointers in the legacy classes' docstrings. No
DeprecationWarning is emitted -- legacy transforms keep working
unchanged, and the as_tool() adapter from the previous commit gives
users a soft migration path.

New tools under torchrl.envs.llm.agentic.tools:

• MCPToolset / MCPServerConfig: connects to an MCP server over stdio
  and materialises remote tools as N native Tool instances, each with
  the server's auto-discovered input_schema. Replaces the legacy
  MCPToolTransform's background-thread plumbing with native asyncio.
  Optional dependency on the 'mcp' package.
• HttpTool: stdlib-only async HTTP via urllib.request offloaded to a
  worker thread. Optional allowed_hosts list plus a hard
  max_response_bytes cap. Pair with RateLimiter under ToolCompose for
  per-host throttling.

torchrl.envs.llm.agentic top-level exports gain HttpTool,
MCPServerConfig, MCPToolset.

Documentation:

• New sphinx-gallery tutorial tutorials/sphinx-tutorials/llm_agentic.py
  walking through a minimal agentic loop (Python + Stop), parser
  swapping (XML / JSON / OpenAI / Anthropic), parallel-dispatch
  motivation, the as_tool() migration recipe, and an MCP example.
  Smoke-tested end-to-end (prints '<tool_result call_id="c1">4</tool_result>').
• docs/source/reference/llms_envs.rst gains a "Built-in tools and
  adapters" autosummary block and a migration table mapping each legacy
  class to its agentic counterpart and adapter recipe.

Soft .. seealso:: pointers added to the docstrings of:
ExecuteToolsInOrder, PythonInterpreter, SimpleToolTransform,
MCPToolTransform, BrowserTransform, XMLBlockParser, JSONCallParser.
Each points readers at the modern equivalent or the as_tool() adapter
without changing runtime behaviour.

Tests extend test/llm/test_llm_transforms.py with TestMCPToolset
(import-gated, server config) and TestHttpTool (host allowlist,
protocol conformance).

Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/3737

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 5 New Failures

As of commit ff12538 with merge base d386287 ():

NEW FAILURES - The following jobs have failed:

• Lint / lint-done (gh)
  Process completed with exit code 1.
• Lint / python-source-and-configs / linux-job (gh)
  tutorials/sphinx-tutorials/llm_agentic.py:102:1: F401 'torchrl.envs.llm.agentic.parsers.OpenAIToolCallParser' imported but unused
• LLM Tests on Linux / unittests-vllm (3.12, 12.9) / linux-job (gh)
  test/llm/test_wrapper.py::TestLogProbsComparison::test_log_probs_consistency[unpadded-tokens]
• Unit-tests on Linux / tests-olddeps (3.10, 11.8) / linux-job (gh)
  test/test_tensordictmodules.py::TestGRUModule::test_gru_scan_prototype
• Unit-tests on Windows / unittests-cpu (3.10, windows.4xlarge, cpu) / windows-job (gh)
  test/test_tensordictmodules.py::TestGRUModule::test_gru_scan_prototype

This comment was automatically generated by Dr. CI and updates every 15 minutes.