Changes from 3.6.0 to 3.6.1

This is a bugfix and performance release that optimizes read-ahead prefetching
for memory-mapped frames on GPFS (IBM Storage Scale), avoiding page faults by
populating mappings with madvise(MADV_POPULATE_READ) on Linux.

• Prefetching on GPFS via memory mapping population: For frames read through
  the mmap I/O backend (BLOSC2_IO_FILESYSTEM_MMAP) on GPFS, read-ahead threads now
  populate the mapping directly using madvise(MADV_POPULATE_READ) (Linux >= 5.14)
  instead of read(). Because GPFS caches read data in its internal pagepool rather
  than the kernel page cache, populating the mapping avoids costly page faults and
  speeds up multi-dimensional slice reads by 1.10x–1.34x (#880).

• Cached filesystem type checks: Filesystem type queries via statx() are now
  cached per thread across the last 4 accessed devices. This avoids redundant
  fstatfs() RPC roundtrips to servers on remote filesystems such as NFS, preventing
  I/O latency regressions on cached patch reads (#880).

• Kernel MADV_POPULATE_READ runtime detection: Added a one-time per-process probe
  to verify kernel support for MADV_POPULATE_READ. When unsupported (Linux < 5.14),
  read-ahead threads safely fall back to reading chunks without issuing superfluous
  page-cache queries via cachestat() (#880).

• New BLOSC_PREFETCH_POPULATE environment variable: Setting
  BLOSC_PREFETCH_POPULATE=1 forces memory-mapped frame prefetching to populate
  mappings via madvise(MADV_POPULATE_READ) across all filesystems, while 0
  disables it completely (#880).

• Fixed test build under Clang on Linux: Defined _GNU_SOURCE in
  tests/b2nd/test_b2nd_prefetch.c so madvise() is exposed when building with Clang
  under -D_XOPEN_SOURCE=600 (#880).

Special thanks to Fabian Isensee for these contributions!