fix: Eliminate port allocation race condition in all affected test files

[maxisbey]

This PR extends the port allocation race condition fix to all test files affected by the same TOCTOU issue, eliminating intermittent test failures when running with pytest-xdist parallel workers.

Motivation and Context

When running tests with pytest-xdist parallel workers, multiple workers would receive the same "free" port from socket.bind(("127.0.0.1", 0)), causing a TOCTOU (Time-of-Check-Time-of-Use) race condition. This manifested as:

• "address already in use" errors when multiple workers tried to bind the same port
• Tests connecting to wrong servers (e.g., expecting one server but getting another)
• Flaky tests in CI that would pass locally but fail intermittently in parallel runs

This builds on the initial fix in tests/server/fastmcp/test_integration.py by applying the same solution to 7 additional test files that had the identical issue.

How Has This Been Tested?

• ✅ Ran full test suite with 14 parallel workers: PYTEST_DISABLE_PLUGIN_AUTOLOAD="" uv run --frozen pytest tests/ -v
• ✅ All 703 tests pass consistently
• ✅ Verified no "address already in use" errors occur
• ✅ Confirmed no test cross-contamination between workers
• ✅ Tested multiple consecutive runs to ensure stability

Each pytest-xdist worker now receives an exclusive port range from 40000-59999 based on its worker ID, eliminating port conflicts entirely.

Breaking Changes

None. This is an internal test infrastructure change with no impact on the public API or user-facing functionality.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update

Checklist

• I have read the MCP Documentation
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have added or updated documentation as needed

Additional context

Files modified:

• tests/server/test_streamable_http_security.py - Updated server_port fixture
• tests/server/test_sse_security.py - Updated server_port fixture
• tests/shared/test_streamable_http.py - Updated 3 port fixtures (basic, json, event server)
• tests/shared/test_sse.py - Updated server_port fixture
• tests/shared/test_ws.py - Updated server_port fixture
• tests/client/test_http_unicode.py - Updated unicode_server_port fixture
• tests/client/test_notification_response.py - Updated non_sdk_server_port fixture
• tests/test_test_helpers.py - Fixed socket cleanup to prevent ResourceWarning

Implementation details:
All affected fixtures now use get_worker_specific_port(worker_id) instead of the race-prone socket.bind(("127.0.0.1", 0)) pattern. The helper function allocates worker-specific port ranges ensuring no overlap between parallel test workers.

[maxisbey]

actually, I don't like this approach anymore.

It's a bit late at night so I'm going to leave it as draft for now, but some other options I was thinking are:

• Adding more port randomisation - although this only just reduces the likelihood of port collisions.
• Using a sub process with a mutex or shared memory or something, have that process hand out port numbers - although this means we have a limited number of ports, so we'd need a way to release them over time
• Using a FileLock to allocate ports per test

Although none of them really are clean IMO. So going to leave this for now