Release the GIL during liberasurecode encode/decode/reconstruct

Holding the GIL for the whole C computation kept callers from overlapping erasure coding with other work or spreading it across cores.

Only the liberasurecode_* calls run without the GIL, never any CPython API. Input buffers are borrowed from immutable bytes the caller still holds.

Only do this on liberasurecode 1.8.0 and newer, which holds a read lock across the operation – and thereby asserts its backends are reentrant on a shared descriptor. Older versions walk their instance list unlocked, so an operation overlapping an ECDriver being created or destroyed would race it; the GIL is what kept those apart.

With the threaded bench in change 991506 this scales 425 -> 790 -> 1484 MB/s at -t 1/2/4, against a flat 414 -> 423 without.

Add a test driving a shared driver from several threads.