Lower Linear Algebra factorizations to StableHLO

Instead of using `stablehlo.custom_call` to call LAPACK or TPU kernels, it would be cool to code the numerical methods on top of other StableHLO ops. This has some nice advantages:

- Shardy could know how to automatically shard the linalg factorizations
- Posibly AD could work on linalg factorizations without new rules (although new rules might be better)
- Interesting to see if there's any performance improvement using the opt passes implemented in this repo
- Portable implementations to platforms whose backend do not yet support some of these methods (e.g. cuSOLVER doesn't support all methods available in the LAPACK reference)

# To do

- [ ] LU
- [ ] QR
- [ ] Eigendecomposition
- [ ] SVD
- [ ] QR iteration (LAPACK's `gesvd`)
- [ ] Divide and conquer (LAPACK's `gesdd`)
- [ ] Jacobi method (LAPACK's `gesjv` and `gesvj`)
- [ ] QR-based dynamically weighted Halley iteration (QDWH, used by Jax)
- https://epubs.siam.org/doi/abs/10.1137/120876605
- https://epubs.siam.org/doi/abs/10.1137/090774999
- [ ] Lanczos algorithm for top-k eigenvalues

The issue names no files, tests, or concrete entry point. Start by narrowing the checklist to one factorization and locating the existing stablehlo.custom_call path for LAPACK or TPU kernels. Done would mean a portable implementation using StableHLO operations, with its sharding, autodiff, portability, and performance implications evaluated.

Written by the indexing model from the issue text.