ucs: portability fixes for non-glibc / Clang environments

[GenericRikka]

UCS: portability fixes for non-glibc/Clang environments

This PR addresses a set of portability issues in the UCS layer that
prevent UCX from building cleanly with Clang and/or non-glibc libc
implementations such as FreeBSD libc or musl. All changes are confined
to UCS and have no effect on the wire protocol or ABI.

Changes

sys.h / log.c — portable hostname buffer size (UCS_HOST_NAME_MAX)

HOST_NAME_MAX is a Linux/glibc extension and is not defined by POSIX.
Introduce UCS_HOST_NAME_MAX derived from _POSIX_HOST_NAME_MAX (the
POSIX-mandated minimum of 255) with a fallback of 256, and use it
unconditionally in the hostname buffer in log.c. The <limits.h> and
<unistd.h> includes required to expose _POSIX_HOST_NAME_MAX are
added to sys.h.

compiler.h — UCS_F_NOOPTIMIZE with Clang

Clang does not support GCC's __attribute__((optimize("O0"))) and
silently ignores it, producing no-op behaviour for UCS_F_NOOPTIMIZE.
Clang provides the equivalent __attribute__((optnone)). Restructure
the macro into a clean three-branch #if/#elif/#else so each
compiler gets the correct attribute.

float8.h — remove dependency on <ieee754.h>

<ieee754.h> defines union ieee754_double using C bitfields to
decompose double-precision floats. This header is a glibc extension
absent from other libc implementations. Replace the bitfield-based
access with portable memcpy-based helpers (ucs_double_to_words /
ucs_words_to_double) that inspect the IEEE 754 bit layout without
undefined behaviour, using FLOAT_WORDS_BIGENDIAN (already defined by
UCX's autoconf) for endianness.

aarch64/cpu.h — explicit <ucs/sys/ptr_arith.h> include

The AArch64 CPU header uses macros from ptr_arith.h but relied on
transitive inclusion. Add the direct include to make the dependency
explicit and guard against breakage if include order changes.

debug.c — explicit <signal.h> and <stdlib.h> includes

debug.c uses signal-related types and stdlib functions without
including the corresponding headers directly. Add explicit includes.

numa.c — replace __CPU_SETSIZE with UCS_CPU_SETSIZE

numa.c used the glibc-internal __CPU_SETSIZE (double-underscore
reserved namespace). UCX already defines UCS_CPU_SETSIZE in
<ucs/type/cpu_set.h>, which is available transitively via
<ucs/sys/sys.h>, already included by numa.c. Remove the
now-redundant fallback block and replace all three use sites with
UCS_CPU_SETSIZE. Using the UCX-internal constant is also semantically
correct: the NUMA CPU cache bound should align with UCX's own CPU set
abstraction, not the platform's reported CPU_SETSIZE.

[tvegas1]

maybe we can simply manually define union ieee754_double when it is missing, and avoid most of the changes?