v0.9.1 2025/11/07: More multi-device support; updated CPU PJRT; dropped static CPU PJRT linking.

@timkaye11

Package cmd/gopjrt_installer:
- Link libcublasLt.so.13 and libcublas.so.13 to the lib subdirectory of the installation directory given.
  Nvidia needs it for some models, but doesn't know how to find it within the provided SDK path.
Package xlabuilder: removed from CI: it requires manual installation of the older library version.
Package pjrt:
- Fixed wrong struct size set for PJRT_Event_Destroy_Args, in #65 (@timkaye11)
- Fixed buffer.Destroy to release the client pointer in the wrapper.
- Added a required runtime.KeepAlive(program) on a CGO call to compile program.
- Added new Shardy support for distributed (across multiple devices) execution.
- Added old SPMD support (see pjrt.Compile().WithSPMD)

Package pjrt:
- Added Buffer.CopyToDevice().
Updated and rebuilt XLA's CPU PJRT and XlaBuilder for new release.
- Static builds of PJRT CPU no longer working (Bazel issues).
- XlaBuilder no longer included – making the binary packages smaller.

v0.9.0 2025/11/06: Multi-device SPMD execution; TPU installation support; Deprecating `xlabuilder`.

gopjrt_installer: Added TPU installation.
Package pjrt:
- Renamed Client.Devices to Client.AllDevices.
- Added Client.NumDevices to get the number of addressable devices.
- Added Compile().WithSPMD for multi-device SPMD execution.
- Default compilation is "portable" (can be run in any device), but it allows for device specific as well.
- Arena pools for powers-of-2 sizes.
Replaced xlabuilder with stablehlo in examples:
- Enabled darwin tests using stablehlo
Updated PJRT C API and protos from XLA sources.

GitHub workflows:
- Renamed to more meaningful names.
- Added darwin_build_cpu_pjrt.yaml: workflow to automatically build the Darwin CPU PJRT plugin.
- Remove unnecessary apt install in linux tests.
Package xlabuilder:
- Added -std=gnu11 to CFLAGS to avoid incompatibility introduced in c23.
- Temporary fix, as xlabuilder is deprecated and will be removed in the future.

Added default "${HOME}/Library/Application Support/GoMLX/PJRT" as a default search path for PJRT plugins in MacOS.
Fixed cuda.go to be linux-only for now, with a safe default for other platforms.
- Removed the requirement for static linking of CUDA PJRT plugin for Darwin.
gopjrt_installer: now also supports darwin/arm64.
- Use -tags=all when building to include all installers (linux/darwin) if cross-platform installing.

New github.com/gomlx/gopjrt/cmd/gopjrt_installer command-line tool to install Gopjrt plugins.
- Doesn't do sudo by itself.
- Doesn't depend on Python for CUDA plugins (automatic dependency traversing; sha256 checking).
- Interactive, self-explained mode.
- Optional version specification (both for Gopjrt and for CUDA PJRT).
- All versions in one tool.
- Faster
- Doesn't download unnecessary jax files.*
Updated README.md, with links to gopjrt_installer for installation.
- Moved xlabuilder documentation under xlabuilder/README.md, and marked it as deprecated.
- Changed Mandelbrot example to use github.com/gomlx/stablehlo instead.

Updated PJRT's pjrt_c_api.h and protos from github.com/openxla/xla, and regenerated wrapper code.
Added cmd/install_cuda13.sh script to install "jax[cuda13]" drivers.
Updated README.md to stard adverstising of xlabuilder deprecation.
Moved pjrt.SuppressAbseilLoggingHack implementation to the "per-platform" files pjrt/dynamiclib_posix.go and
pjrt/dynamiclib_darwin.go: the first using syscall.Dup3() and the second using syscall.Dup2().

Moved stablehlo to its own repository.
Updated dependencies.
Package dtypes:
- Added DType.Bits field indicating the number of bits in the data type -- supporting sub-byte types including
  quantized types.
- Added DType.IsPromotableTo(target) to check if types are promotable to the target type -- it follows the StableHLO rule.
- Cleaned up generators for dtypes package.
Package xlabuilder:
- Deprecated SelectAndScatterSum.

API change to convolutions:
- ConvGeneralDilated() renamed to ConvGeneral(). Fixed (standardized) name of the parameters,
  but they remain the same.
- The ConvolveAxesConfig structure had its field names slightly changed, by replacing "Channel"
  by "Feature".
Updated XLA protos.