fix(mcp): pass the session read timeout in the type the installed mcp SDK expects

[chrisnestrud]

mcp SDK 2.x types ClientSession.read_timeout_seconds as a float number of
seconds and adds it to other floats internally, so the timedelta that 1.x
expects made every HTTP MCP server fail to initialize with:

unsupported operand type(s) for +: 'float' and 'datetime.timedelta'

cecli declares mcp>=1.24.0 with no upper bound, so installing 2.x silently
breaks the transport, while the other 2.x shims in the same module
(_get_http_client_module, _unpack_transport, the OAuth callback handler)
kept working.

This routes the value through a helper that returns float seconds for mcp 2.x
and a timedelta for 1.x, so both majors keep the same per-request timeout
instead of failing to connect.

Fixes #683.

Repro, with mcp 2.2.0 installed: cecli --agent logs Error initializing server context7: unsupported operand type(s) for +: 'float' and 'datetime.timedelta' three times per server, then MCP tool initialization failed. After the change, an agent session calls context7
resolve-library-id and deepwiki read_wiki_structure normally, with the
servers configured as website/docs/config/mcp.md documents.

Tests: tests/mcp/test_server_sdk_compat.py gains one case per major;
tests/mcp is 119 passed.