fix(mcp): cache empty tools lists

[CoralGarden52]

This pull request fixes MCP tools-list caching when a server returns a valid empty tool list.

With cache_tools_list=True, MCPServer.list_tools() stores an empty result as [], but the cache-hit check used list truthiness. As a result, every later discovery call fetched the empty list again. The cache-hit check now distinguishes the unpopulated None sentinel from a populated empty list.

The regression test exercises the public MCPServerStdio context-manager and list_tools() boundary, verifies cached_tools == [], and confirms that two calls make only one underlying MCP request. Existing invalidation, pagination, filtering, mutation-isolation, and non-empty caching behavior remains covered by the same test module.

Validation:

• make format
• make lint
• make typecheck
• make tests

All repository verification commands passed locally.

[sylvesterkaczmarek]

I traced this through the cache sentinel and invalidation paths. _tools_list is None is the unpopulated state, while [] is a valid server result, so switching the hit check away from list truthiness fixes the repeated-fetch case without changing the dirty-cache contract: invalidate_tools_cache() still forces the next lookup through the MCP request path. The cached value also remains the raw tool list before filtering, so this does not make a filtered view the new source of truth.

The regression exercises the public context-manager/list_tools() boundary, confirms cached_tools == [], and pins a single underlying request across two lookups. I did not find a blocking issue in this change.