perf: pool http connections per miner scoring iteration

[seroperson]

Validator scoring makes many GitHub REST and GraphQL requests per miner and each of those previously opened a new TCP socket and negotiated a new TLS session. This PR introduces a requests.Session-based pool scoped to one miner's fetch iteration, so every http call within that iteration reuses the same kept-alive TCP connection.

Re-using existing connections makes things bit faster and overall it's a good practice when you connect to the same host again and again.

Summary

• New session_scope() context manager holds a module-level _session_cache: Optional[Dict[str, requests.Session]], populated on enter and cleaned up (sessions .close()d, cache set back to None) on exit.
• New get_session(token) returns a per-PAT cached Session inside a scope and a fresh uncached session outside it. Tokenless calls route through the same helper via get_session(token or '').
• New make_anonymous_headers() mirrors the existing make_headers / make_graphql_headers pattern for unauthenticated GitHub calls.
• All call sites now use session = get_session(token)
• session_scope() wraps load_miners_prs + score_miner_prs inside evaluate_miners_pull_requests so one cache covers an entire miner's load+score cycle; connections close before the function returns.

Testing

No new tests are introduced. Added the tests/conftest.py, which installs an autouse _ForwardingSession shim that delegates .get/.post back to the module-level requests.get/post, so existing dozens of tests which use @patch('requests.get'/'requests.post') keep working unchanged.

[anderdc]

Can you prove the actual performance gain from this change? Otherwise it seems like changing something that already is proven to work

[seroperson]

Can you prove the actual performance gain from this change? Otherwise it seems like changing something that already is proven to work

Not a game-changing x10 of course, but still something. Considering how many requests it does during a round, it should result into noticeable gain.

$ > GITHUB_PAT=... python /tmp/bench_session_scope.py

N=15 sequential authenticated /user calls, 3 repetitions
==============================================================================
 run 1:
  fresh Session per call        6306.4 ms  (420.4 ms/call)  TCP conns: 15
  one shared Session            4796.4 ms  (319.8 ms/call)  TCP conns: 1
 run 2:
  fresh Session per call        6106.1 ms  (407.1 ms/call)  TCP conns: 15
  one shared Session            4481.7 ms  (298.8 ms/call)  TCP conns: 1
 run 3:
  fresh Session per call        7436.8 ms  (495.8 ms/call)  TCP conns: 15
  one shared Session            4672.7 ms  (311.5 ms/call)  TCP conns: 1

bench_session_scope.py

[seroperson]

Updated to match current test. I see MirrorClient also uses sessions, so now it has bit more value to get in line =)

[anderdc]

Bench is convincing — happy to take this. Three asks before merge:

1. Mirror parity. MirrorClient already holds a requests.Session (gittensor/utils/mirror/client.py:44), but in evaluate_miners_pull_requests (reward.py:75-83) load_mirror_miner_prs and score_mirror_miner_prs each default client = client or MirrorClient(), so per miner there are two separate Session pools to mirror.gittensor.io and no connection reuse across the load→score boundary — the same waste this PR fixes for the PAT path. Instantiate one MirrorClient() in evaluate_miners_pull_requests and thread it through both calls via the existing client= kwarg.

2. Accept-header bleed into GraphQL. _build_session applies make_headers(token) to session.headers, which sets Accept: application/vnd.github.v3+json. In execute_graphql_query and get_github_graphql_query the per-request headers=make_graphql_headers(token) only overrides Authorization/Content-Type, so GraphQL POSTs now carry the v3 REST Accept they didn't before. GitHub tolerates it, but it's an unintended drift. Either drop Accept from _build_session (let call sites own it) or add Accept: application/json to make_graphql_headers.

3. Cover `session_scope` directly. The new `tests/conftest.py` shim swaps `get_session` autouse, so the suite doesn't exercise `_session_cache`, the re-entrance guard, or close-on-exit. Add a small test: same Session returned for same token within a scope, distinct per token, re-entrance raises, sessions `.close()`d on exit. While there: the shim calls `make_headers(token)` unconditionally, but prod uses `make_anonymous_headers()` for `token=''` — minor test/prod drift on the unauth REST search fallback.