Release slab memory to the OS between iterations

[chourroutm]

Copying/downsampling processed slabs with steadily growing resident memory: the large numpy copy buffers and tensorstore shard-encode buffers were freed at the Python level but retained in the glibc allocator arena rather than returned to the OS, so RSS ratcheted up slab-by-slab until a memory-capped session was exhausted.

• Add _release_memory() (gc.collect + guarded glibc malloc_trim) and call it at the end of _copy_slab and _downsample_block after dropping the slab/shard buffers and tensorstore handles.
• Bound the tensorstore cache pool so the downsample read path cannot grow an unbounded in-memory cache.
• Build the joblib job list lazily so each slab's dask view is released as it completes instead of all being pinned for the whole call.
• Document that peak memory scales with n_processes.