Update jax-ml/jax to commit 6aaf234b1ee98800fcb80b9014a052a66dadbd8a - #3368

Rebase the patches onto the XLA (ed6af492) and LLVM (b08e0059) it pins:
- xla.patch: the rocm BUILD.tpl hunk takes its new context; the
  topology_util.cc hunk (#3376) is upstream and dropped.
- llvm_amdgpu_bb_prolog.patch: new context only; llvm/llvm-project#225433
  is still open.
- llvm_mlir_import_inrange_width.patch (#3367) and
  llvm_orc_unw_revert.patch are upstream and no longer applied.
- The Windows guard appended to stablehlo's temporary.patch is anchored
  where XLA's own InterpreterOps.cpp change now leaves evalRunParallelOp.

And build against that LLVM:
- vendored Polly: IRInserter::InsertHelper no longer overrides, as the
  IRBuilder inserter is no longer virtual (llvm/llvm-project#229700).
- populateGpuToROCDLConversionPatterns takes a ROCDL::TargetInfo in place
  of an amdgpu::Chipset (llvm/llvm-project#223562); the default one is
  the same unknown target.
- raiseimmerse3: the new arith fold of or(x, or(y, x)) changes the
  order affine-cfg reaches the compare in, which is now left as an
  affine.if of true/false rather than turned into a cmpi eq; the same
  happens on main when the input is given that fold already.
- CompareOpCanon: XLA's stablehlo patch gives the reference compareOp
  the op's optional compare type; pass it through.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016zErYp7upmqr4NHfhod9UD