Guard NextActingInFixedOrder's turn state with its own lock

[jaisinha77777]

Summary

While auditing this codebase's lock usage after the `Measurements.close()` deadlock fix (#282) and the `CallLimitLanguageModel` counter fix (#285), I found `NextActingInFixedOrder` (in `concordia/components/game_master/next_acting.py`) is inconsistent with its own locking discipline:

• `remove_actor_from_sequence()`, `add_actor_to_sequence()`, `get_state()`, and `set_state()` all take `self._lock` before touching `self._sequence`/`self._currently_active_player_idx`.
• `pre_act()` and `get_currently_active_player()` did not, despite reading and mutating exactly that same state.

Concretely: `pre_act()` reads `len(self._sequence)` and then indexes into `self._sequence`, while `remove_actor_from_sequence()` (which does take the lock) mutates that same list. If these ever run on different threads at the same time, `pre_act()` can compute an index against the old length and then index into a list that's since been shortened.

Being precise about confidence here, same as in #285: within a single `EntityAgent.act()` call, phases are sequenced and (from what I could trace) only one component's `action_spec.output_type` condition is normally true at a time, so I could not identify a call path where this is reachable today, and I could not force an `IndexError` out of it even under heavy synthetic thread contention (many threads racing `pre_act`/`get_currently_active_player` against `remove_actor_from_sequence`/`add_actor_to_sequence`). So I'm not claiming this is an observed live bug. What I can say confidently: this class was the one inconsistent case, both among its own methods and relative to every sibling class using this same locking pattern elsewhere in the codebase (`Measurements`, `ObservationQueue` in this same file, etc.), that left this particular state unprotected.

Fix

Takes `self._lock` in `pre_act()` (only around the read-index-and-mutate block, not the whole method) and in `get_currently_active_player()`, matching the rest of the class.

Test plan

New `concordia/components/game_master/next_acting_test.py` (previously no coverage at all):

• Sequence cycling and wraparound order.
• Non-`NEXT_ACTING` action specs are ignored.
• `remove_actor_from_sequence`/`add_actor_to_sequence`, including the `ValueError` case for an unknown actor.
• `get_state`/`set_state` round-tripping, including that a restored component continues the original sequence correctly.
• A concurrency test driving `pre_act`/`get_currently_active_player` against `remove_actor_from_sequence`/`add_actor_to_sequence` from multiple threads and asserting no exception is raised.

`python -m pytest concordia/components/game_master/ -q` → 177 passed (full directory, no regressions).
`python -m pyink --check` on the new test file → clean. (`next_acting.py` itself has substantial pre-existing formatting drift unrelated to this change in classes I didn't touch; I left that alone to keep this diff focused.)