feat: default --xla_gpu_autotune_level=0 in jax_wrapper (env-respecting)

[Jammy2211]

Overview

Rollout of autolens_profiling#74 verdict 2: the pathological ~7m30 cold GPU compile is XLA autotuning. Evidence (A100): autotune genuinely off gives a 17× cold-probe reduction (498 s → 29 s), −40 % cold full fit (2081 s → 1253 s), bit-identical fixed-input logL, and steady-eval parity across the measured matrix (worst ~4 % on one cell; the 4800-eval fit was faster end-to-end). Stacked with the default cache (#127/#128), worst-case first-fit UX drops ~70 min → ~30 s.

Plan

• autoconf/jax_wrapper.py: append --xla_gpu_autotune_level=0 to the assembled XLA_FLAGS unless the user's flags already contain an --xla_gpu_autotune_level setting (any value — respect it). Same append/env-respecting pattern as feat: JAX persistent compilation cache by default + XLA_FLAGS append fix (#127) #128. INFO log line.
• Extend test_autoconf/test_jax_wrapper.py: flag added when absent; user-preset level respected; composes with the constant_folding append.
• Ship validation: fresh-process check that the flag reaches XLA_FLAGS and a preset level survives.
• Downstream PyAutoFit suite against the branch (base-layer change).

Prompt: PyAutoMind draft/feature/autoconf/default_xla_gpu_autotuning_off_in_the.md · branch feature/autotune-off-default

[Jammy2211]

Merged via PR #132 — autotuning off by default (env-respecting). With #128's cache: worst-case first fit ~70 min → ~30 s. Closing.