Reduce redundant host work in the xFit LM solver

[tpn]

The normal-equation path copied five arrays even when every row was valid,
and filtered trial-evaluation budgets after analytic Jacobians that consume no
residual evaluations. Skip those operations to reduce repeated host work in
xFit's shared solver. Preserve fallback filtering and finite-difference budget
checks, with a regression for exhausting the budget just before a trial in both
implicit and explicitly requested finite-difference modes. Builds on #85.

Validation on the integrated source (Python 3.12, CUDA 13):

• 712 CPU tests passed, including solver budget, shared execution, and pipeline coverage.
• All 70 native and Numba scientific, CLI, and end-to-end pipeline tests passed without skips.
• All-file hooks passed with the optional GPU and Numba dependencies installed, including mypy and lock validation.

Earlier free-threaded Python 3.14 testing with Numba-CUDA-MLIR 0.5.4 exposed
an intermittent concurrent-startup failure in dependency type dispatch.
These Python 3.12 runs do not qualify free-threaded concurrent-startup reliability.

[coderabbitai]

Important

Review skipped

Auto reviews are limited based on label configuration.

🏷️ Required labels (at least one): 1

• ai-review

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cuPhoton/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: eccb0e59-7700-4e2a-bc4c-d8716567134d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}