Precondition midpoint Hermite streaming lines

[rogeriojorge]

Summary

• add an opt-in implicit-midpoint preconditioner for collision, diffusion, and x-directed Hermite streaming
• use SOLVAX's existing JAX-native banded LU rather than adding a SPECTRAX linear solver
• keep electromagnetic fields unchanged and treat Hermite-shard boundaries as outer-GMRES block couplings
• leave every solver default and tolerance unchanged

Motivation and evidence

The existing diagonal preconditioner removes collision/diffusion stiffness but not the tridiagonal free-streaming coupling in Hermite space. This change solves that physics block exactly along each local x-Hermite line.

On the four-device 513-Fourier by 256-Hermite Landau run to t=5, it reduced trajectory time from 196.015 s to 21.521 s and GMRES work from 46,015 (maximum 119 per step) to 3,144 (maximum 9). Relative energy change remained 3.83e-13; endpoint coefficient difference was 1.06e-12 relative and the final/initial field-norm ratio changed from 0.4358742732 to 0.4358742563.

On 257 by 128 Landau to t=1, runtime fell from 8.131 s to 3.462 s and GMRES work from 1,960 to 231.

This is deliberately not a replacement for the diagonal preconditioner. On 2-D 65-squared Orszag--Tang, iteration counts were unchanged and line-factorization overhead increased runtime from 5.776 s to 6.849 s. A two-stream control also confirms that callers must retain accuracy-appropriate linear tolerances; no tolerance is changed here.

Verification

• 9 tests pass on the complete midpoint stack
• Ruff and git diff checks pass
• isolated source and wheel builds pass
• the new unit test compares the complex banded solve directly with a dense midpoint operator containing drift, collision, diffusion, and both Hermite off-diagonals
• combined species/Hermite-sharded audit runs complete with global Newton/GMRES products

Scope

This PR intentionally does not cache factors across time steps, add y/z alternating-direction solves, change midpoint defaults, or modify SOLVAX. Those extensions should be benchmarked independently.