fix(sampling): escalate temperature at the exact retry midpoint

[Osamaali313]

Summary

dynamically_adjust_temperature (in concordia/utils/sampling.py) returns 0.0 when attempts == max_attempts / 2, which is reachable whenever max_attempts is even.

temperature = 0.0
if attempts > 1 and attempts < (max_attempts / 2.0):
    temperature = 0.5
elif attempts > (max_attempts / 2.0):
    temperature = 0.75
return temperature

Both branches are strictly exclusive (< max/2 and > max/2), so the exact midpoint falls through to the default 0.0. For example with max_attempts=10, attempts=5 returns 0.0 — dropping the choice-sampling temperature back to zero for that retry instead of escalating it, which defeats the retry-diversity intent at exactly the point where it should be ramping up.

Fix

Make the upper branch inclusive so the midpoint escalates:

elif attempts >= (max_attempts / 2.0):
    temperature = 0.75

• attempts=5, max=10 now returns 0.75 (was 0.0).
• All other attempt counts are unchanged (verified for every attempts over max_attempts in 2..20).
• Odd max_attempts (no integer midpoint) are unaffected.

Tests

Adds concordia/utils/sampling_test.py — the first unit tests for this module — covering the temperature schedule (incl. the midpoint regression) and extract_choice_response. All 11 tests pass.

[Osamaali313]

Heads-up on the red Test Concordia run: the failure is in the Typecheck (pytype) step, on files this PR does not touch —
concordia/components/game_master/interrupt_resolution_test.py and interrupt_scheduling_test.py
(No attribute 'expiry' / 'description' / 'entity_name' on None). Both files exist on main, so this is a pre-existing typecheck failure rather than something introduced here; pytype runs repo-wide and ninja stops on the first error, so the whole job goes red.

This PR only changes concordia/utils/sampling.py (the one-line midpoint fix) and adds concordia/utils/sampling_test.py. Both typecheck cleanly in the same run, and the new tests pass. I'd be glad to send a separate PR adding the missing None-narrowing in those interrupt_*_test.py files if that would help get the typecheck green again.