deps: move the GPU stack to CUDA 13 + numpy 2, and keep configure_jax working across the JAX 0.10/0.11 config rename

[geekbeast]

Two commits that together move the GPU stack to CUDA 13 and keep configure_jax working on the JAX releases that stack pulls in.

1. Move the GPU stack to CUDA 13 and numpy 2 (c511158)

requirements.txt asked for cupy-cuda12x and jax[cuda12] while pyproject.toml's jax-gpu and all extras already asked for jax[cuda13] — the two files disagreed about which CUDA major this project targets. Both cupy wheels provide the same cupy module, so installing across the two shadows one rather than raising a conflict.

The numpy pin is the part that quietly costs a GPU. jax only gained CUDA 13 plugins in 0.10, and jax>=0.10 requires numpy>=2.0 — so numpy==1.26.4 forces the resolver down to a jax with no CUDA 13 plugin at all. It installs, imports, and runs on CPU next to a GPU-enabled torch and cupy, warning only in a log line. Relaxing to numpy>=2.1 pulls in jax-cuda13-plugin / jax-cuda13-pjrt. The <2.3 ceiling is numba 0.61.2's, so that pin stays.

• numpy==1.26.4 → numpy>=2.1,<2.3
• cupy-cuda12x==13.5.1 → cupy-cuda13x>=13.6.0
• jax[cuda12] → jax[cuda13]>=0.10.0
• torch>=2.2.2 → torch>=2.5.0, matching the floor pyproject already declared (CUDA 13 additionally needs a cu130 build, which current PyPI wheels provide)
• pyproject's all extra combined jax[cuda13] with cupy-cuda12x, which cannot both be right — now cupy-cuda13x. The existing cupy extra stays on cuda12x for anyone still on that toolkit, with a new cupy-cuda13 extra alongside.
• Comment in requirements-dev.txt noting sigkernel's setup.py imports Cython without declaring it in [build-system].requires, so it needs --no-build-isolation.

Dependency metadata only — no source changes.

2. Keep configure_jax working across the JAX 0.10/0.11 config rename (01f9deb)

JAX 0.11 removed jax_exec_time_optimization_effort and jax_memory_fitting_effort, replacing the continuous 0.0–1.0 floats with the jax_optimization_level / jax_memory_fitting_level enums (UNKNOWN/O0–O3). jax.config.update raises AttributeError on an unrecognised key, so on 0.11 configure_jax died partway through — after enabling x64 and setting the XLA flags, but before matmul precision and the compilation cache were ever configured.

Pinning jax<0.11 isn't an option: this project declares jax>=0.10.0 and 0.10 only accepts the old spelling, so a pin would trade one broken half of the declared range for the other. _update_first_supported applies the first key the installed JAX recognises, leaving both ends working.

Values carry the old intent over: 1.0 was maximum execution-time effort → O3. Memory fitting was deliberately 0.3 for high-memory machines and the new default is O2, so O1 keeps it below default.

Verification

Per the commit messages, verified on CUDA 13.0 (driver 580.178.04, 4x GPU): torch, cupy-cuda13x, and jax 0.11.1 all report GPU; numba 0.61.2 compiles against numpy 2.2.6. On 4x H100 with jax 0.11.1, configure_jax reports O3/O1 with x64 live and GPU detected, and tests/test_core_jax.py passes 21/21.

CI coverage note: the CI workflow installs .[jax-cpu,dev], and the jax-cpu extra (jax[cpu]>=0.4.34) is untouched here — so CI exercises the configure_jax compat fix against whatever JAX the CPU extra resolves to today, but does not cover the CUDA 13 / cupy / numpy changes, which live in requirements.txt and the GPU extras. Those rest on the local 4x GPU verification above.