perf: Replace dict-based request ID management with list-indexed structure

[mykaul]

Summary

The per-connection request ID management uses deque(range(N)) for the free ID pool and dict{int → (cb, decoder, result_metadata)} for in-flight request tracking. Replacing these with list-based structures yields measurable per-request savings and significant memory reduction.

Current Architecture (connection.py)

Structure	Type	Purpose
request_ids	deque(range(300))	Pool of available stream IDs
_requests	dict{int → tuple}	Maps in-flight stream IDs to callbacks
orphaned_request_ids	set()	Timed-out stream IDs awaiting late responses

Proposed Change

• request_ids: deque → list (used as stack with pop()/append())
• _requests: dict → list[tuple|None] (indexed by stream ID, None = free)
• orphaned_request_ids: keep as set() (rarely used)

Benchmark Results (CPython 3.14)

Full request cycle (get ID + store request + retrieve request + return ID):

Approach	ns/op	Memory (300 IDs)
Current (deque + dict)	72.3 ns	~20.6 KB
Proposed (list + list)	47.1 ns	~4.9 KB
Saving	25.2 ns (35%)	~15.7 KB (76%)

Per-operation breakdown (dict vs list for _requests):

Operation	dict	list	Saving
Store request	43.7 ns	13.1 ns	30.1 ns
Retrieve request	43.2 ns	13.1 ns	30.1 ns

Key Implementation Concerns

1. error_all_requests() (connection.py:1143): Currently does requests = self._requests; self._requests = {} (atomic swap). With a list, this becomes swap + allocate new [None]*size, or iterate+clear.
2. _requests.pop(stream_id) with KeyError (connection.py:1407, cluster.py:4509): Needs conversion to if lst[stream_id] is None check.
3. requests.popitem()[1] (connection.py:1162): Needs to find a non-None entry — requires iteration or count tracking.
4. not self._requests truthiness check (asyncorereactor.py:458): Needs a separate _requests_count tracker or any() call.
5. Dynamic sizing: Start at 300, grow list when highest_request_id exceeds current size (matching existing deque growth pattern).

Files to Modify

• cassandra/connection.py (~15 lines)
• cassandra/cluster.py (~3 lines)
• cassandra/io/asyncorereactor.py (~1 line)
• Tests that inspect _requests as a dict

Impact

• ~25 ns saved per request (every read and write)
• ~15.7 KB memory saved per connection (76% reduction)
• Similar magnitude to PR perf: replace RLock with Lock where re-entrant locking is not needed (~11ns saving, -14%) #796 (RLock→Lock, 11ns) and PR perf: lazy-init _callbacks/_errbacks in ResponseFuture (25ns saving, 2x speedup, 112 bytes saved) #795 (lazy callbacks, 25ns)

Related: #536