fix: detect install method when generating .mcp.json

[AR6420]

Fixes #75

Problem

The install command hardcodes "uvx" as the command in the generated .mcp.json.
Users who install via pip don't have uvx on PATH, so the MCP server fails to connect in Claude Code.

Solution

Added a shutil.which("uvx") check when generating the MCP config.
If uvx is available, the existing behavior is preserved.
If not, it falls back to using "command": "code-review-graph" directly, which works regardless of install method.

Testing

• Existing tests pass
• Manually verified: code-review-graph serve works with the fallback config

[JiwaniZakir]

The detection logic in _build_server_entry uses shutil.which("uvx") at call time, which is a reasonable approach, but it's worth noting this check runs on the machine generating the .mcp.json, not on the machine that will ultimately execute it. If a developer generates the config on a system where uvx happens to be installed but deploys it to an environment where it isn't (or vice versa), the generated entry will be wrong. Consider documenting this assumption or accepting an explicit install_method parameter so callers can override the detected value.

Additionally, the entry: dict[str, Any] type annotation is repeated in both branches of the if/else, which is redundant — declaring it once after the block (or using a single assignment with a ternary-style approach) would be cleaner and avoid the Pylance/mypy warning about re-declaring a variable in the same scope.

It's also worth adding a test that mocks shutil.which to return None and verifies the fallback "code-review-graph" command is used, since that path currently has no coverage implied by the diff.

[AR6420]

Quick clarification on the cross-machine point.
"code-review-graph install" is a Local dev tool setup command. It writes .mcp.json on the same machine where Claude Code reads it. There's no generate-here-deploy-there workflow for config files.

Claude Code needs this file locally to connect to the MCP server. So the shutil.which check at install time is the right time to check, because that's the machine that'll use it.

Also worth noting: the current code hardcodes uvx unconditionally so the "wrong machine" concern applies even more to main than to this PR. Good catch on the type annotation and test coverage though, I'll clean those up later.