Support sycl backend for Intel GPU

[khanhnd61-vr]

Adds an Intel GPU backend through oneAPI SYCL, and collapses the per-arch
backend selection that made adding one painful.

src/backend.h is the new shared ladder. Every arch used to open-code the same
accelerator-then-CPU sequence, so a new backend meant editing a dozen files;
they now call vla::backend_init and the ladder lives in one place. Net -225
lines across src/models/. Exactly one accelerator is compiled in, picked by
the CMake flag (GGML_CUDA / GGML_SYCL / GGML_METAL), and configure now
fails if two are on rather than handing ggml both backends. BitVLA keeps its own
CPU init on purpose: it pins its ggml graph to the CPU and offloads through its
own ternary CUDA kernels.

VLA_DEVICE selects the ordinal on both multi-device backends. The SYCL path
range-checks it, because ggml_backend_sycl_init guards the index with an
assert that a Release build compiles out.

Results

Intel Arc A380 (6 GB, Xe-HPG) against the CPU backend on an AMD Ryzen 5 5500,
8 threads. Best of 5-10 iterations after 3 warmups, fixed noise.

Model	input	CPU	Arc A380	speedup
SmolVLA	512	1,920 ms	630 ms	3.0x
Evo-1	448	7,695 ms	1,176 ms	6.5x
VLA-Adapter	224	2,994 ms	517 ms	5.8x

Checked against the CPU backend on every model: max absolute deviation 2.9e-3 on
actions peaking at 0.99, RMS 2.4e-4. That is BF16/F32 kernel rounding, not a
numerical regression.

Two SYCL issues worked around

GGML_SYCL_ENABLE_VMM=0 is defaulted at backend init. ggml-sycl's VMM pool
hands out pointers oneDNN cannot wrap in a dnnl::memory, which aborts the GEMM
for any non-F32 src0 -- that is most checkpoints. It is only a default;
setting it explicitly wins. Disabling oneDNN also clears the crash but costs
~8%.

ggml-sycl has no bf16 -> f32 copy, so VLA-Adapter needs
VLA_ADAPTER_F32_WEIGHTS=1. SmolVLA and Evo-1 run on their stock BF16 defaults.

Setup, the compute-runtime prerequisites and the memory ceiling are in
docs/backend/sycl.md.