Call MLX linalg ops directly on batched inputs instead of mx.vmap

[jessegrabowski]

funcify_Blockwise on MLX now calls a core function directly when that function sets natively_batched = True, after broadcasting the inputs to the common batch shape. Before, every Blockwise went through nested mx.vmap. Every mx.linalg dispatch sets the flag, the same way numba's core functions set handles_out. mx.vmap has no rule for the LUF and QRF primitives, so batched solve, lu, lu_factor, qr, det and slogdet raised. Its rule for solve_triangular calls linalg::inv with the triangular flags dropped, so a batched solve_triangular or cho_solve returned the full solve with nothing raised.

The solve dispatches expand a vector b to a column and squeeze it back. mx.linalg.solve only special-cases a 1-d b, and a batched vector rhs is 2-d.

PivotToPermutations stays on mx.vmap. It is a hand-written scatter loop with no mx.linalg equivalent.

The two test_mlx_eigh failures on this branch are the MLX 0.32.2 eigenvector sign flip that #2424 fixes.

Closes #2385. Supersedes and closes #2399.

[ricardoV94]

Can we open an issue with them to implement the vmap rules for these on MLX. Silly to have to patch around it.

[ricardoV94]

Add other PR as co-author?