Add a device-side allocator

[maleadt]

Implement GPUCompiler's malloc hook so Julia objects that survive optimization can be
allocated in oneAPI kernels. This includes boxed values, Refs passed to non-inlined
functions, and allocations on exception paths.

The motivating failure is GPUCompiler #906:
preserving inferred invoke specializations exposed allocations inside DomainError
constructors, causing kernels such as sqrt.(::oneArray{ComplexF32}) to fail compilation
with an unresolved gpu_malloc. GPUCompiler #908
now proposes a null-returning fallback for backends without an allocator. This PR supplies
actual storage, so surviving allocations can execute instead of always taking the OOM path.

Each work-item gets a 1 KiB private arena. Allocations advance a cursor in 16-byte increments;
nothing is freed individually, and the arena is reclaimed when the work-item exits.
Exhaustion prints ERROR: Out of dynamic GPU memory (trying to allocate N bytes) and
terminates that work-item. It does not raise a host-side exception, so output may be incomplete.
Objects in the arena must not be shared between work-items or retained across launches.

Private memory matches the pointers Julia emits for boxed objects: address space 0 maps
to SPIR-V private memory. A USM buffer cannot back those pointers on Intel GPUs; the
earlier investigation in #621 reported
lost stores through such pointers on a Max 1550.

GPUCompiler threads a hidden KernelState argument through device code. Before SPIR-V
argument lowering, the compiler reserves the arena in the kernel entry block, initializes
its cursor and capacity, and puts its pointer in the state. The pass looks for reads of
the heap field, so it also handles inlined allocator calls. Kernels without those reads
reserve no arena. Device compilation may optimize away heap storage, but there is no
general guarantee that allocations are free.

This extracts the allocator approach from #621 without its exception mailbox, constructor
overrides, or broader compilation tests. Host-visible exception reporting remains separate.
The existing error-printing overrides remain useful because ordinary device exceptions
still do not report their reason to the host.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[maleadt]

@michel2323 I figured that landing an allocator first is more important, instead of bundling several features in #621.

[michel2323]

Thank you! I'm OOO until next week, but I'll try to read it before.

I'm glad you're taking a look. Obviously, I heavily used Claude and I did that bundled PR along benchmarking a code.