Raise the urllib3 pool maxsize to 10 and make it configurable

[jakobovski-arb]

Problem

BaseClient builds its PoolManager with num_pools but never maxsize, so urllib3's default of one connection per host applies:

self.client = urllib3.PoolManager(
    num_pools=num_pools,
    headers=self.headers,
    ...
)

num_pools caps the number of distinct host pools, which doesn't help a client that talks to a single host from many threads. The two are easy to conflate — passing a large num_pools looks like it should raise concurrency, but it changes nothing for a one-host client.

Under threads the result is a TLS handshake storm: threads that lose the race for the single pooled connection open their own, then can't return it (Connection pool is full, discarding connection), so it's closed and the next request renegotiates TLS.

Measurements

Against the trades endpoint, 240 calls, same machine:

	CPU per call	pool is full warnings
single-threaded	0.0074 s	0
20 threads	0.1309 s (18x)	70

And 40 threads were slower than 20 — the pool was the bottleneck, not the network. With maxsize raised: warnings gone, CPU per call down 2.9x, and throughput scales with threads again (25.7 → 66.3 req/sec in that benchmark).

Change

Adds a maxsize parameter to RESTClient, threaded through to the PoolManager, defaulting to 10.

The new default intentionally changes behaviour for every user, including those who never touch the parameter — that is the point of the change. maxsize=1 isn't a sensible ceiling for a client whose own constructor advertises pooling, and a user hitting it has no way to discover why their threads are slow: the symptom looks like a slow API, not a client setting. Fixing it only for people who go looking for a new argument would leave the common case broken.

10 matches both the existing num_pools default and requests' HTTPAdapter default, so it's a conventional value rather than a tuned one.

maxsize is a cap, not a preallocation — a single-threaded client still opens exactly one connection and sees no difference. The change is felt only by concurrent callers, who are the ones currently paying for it.

On BaseClient it's an optional keyword rather than a required one, so existing subclasses calling super().__init__(...) keep working.

Five lines of production change; tests cover the default, an explicit value, and that it's independent of num_pools.

Checks

• pytest test_rest test_websocket — 44 passed, no new failures. (test_clint_headers_concat already fails on a clean master for me, unrelated to this change.)
• black 24.8.0 --check clean on all three touched files.
• mypy clean on both changed modules.