The write path was allocating a fresh buffer with make() for
every file written, then relying on GC to clean it up. The
read path already uses a buffer pool (via AsyncReader) for
reuse and memory tracking.

Use pool.Global() for write buffers too, so they are reused
across transfers and tracked by the memory limiter.

Signed-off-by: Patrick Farrell <pfarrell@whamcloud.com>

Add --local-io-size (default 1M), --local-read-io-size, and
--local-write-io-size flags to control the I/O block size
used by the local backend. Values must be 4K-aligned.

The write path uses pool.GlobalWithSize() to get a properly
sized buffer from a per-size pool, so custom sizes still get
buffer reuse and memory tracking.

The pool gains GlobalWithSize() which returns the default
global pool for the default 1MB size, or creates and caches
a new pool for other sizes.

The AsyncReader gains a NewSize() constructor that uses a
matching pool, so custom read buffer sizes also get proper
pool integration.

Fixes: rclone#9281

Signed-off-by: Patrick Farrell <pfarrell@whamcloud.com>