[zcbenz]

Inspired by jax.config, add a new config module that allows overriding configuration options at run time, as a more relaxable alternative to environment variables.

An example is to disable TF32 for certain ops instead of the whole program:

mx.config.update("MLX_ENABLE_TF32", 0)
high_precise_c = a @ b
mx.config.update("MLX_ENABLE_TF32", 1)

There is also a syntax sugar:

with mx.config.scoped_update(MLX_ENABLE_TF32=0):
  high_precise_c = a @ b

Another example is #4562, in which users want to tune MLX_MAX_MB_PER_BUFFER only for a partial graph instead of the whole program.

Closes #4521.
Closes #4595.

This PR also removes env::get_var and replaces all calls with config::get, which is aware of the configs set at run time.

AI usage disclosure: C++ tests

[aidiffuser]

Tested this as the replacement for #4562 (command-buffer limits raised around decode steps only), on one M3 Ultra (512 GB) with stock mlx-lm main.

This PR (7596fca) against its base (9702959). Greedy decoding, 256 tokens, the default 2048-token prefill chunks. Decode tok/s at short / 32k context; percentages are against the mean of four defaults runs (two per build). Generated tokens were identical on every run.

	MiMo-V2.6-Flash	GLM-5.3-Flash 6-bit
base, defaults	46.2 / 41.7	22.3 / 19.9
this PR, defaults	46.4 / 41.6	22.4 / 19.9
scoped_update(MLX_MAX_OPS_PER_BUFFER=150, MLX_MAX_MB_PER_BUFFER=5000) around decode	+5.6% / −1.6%	+4.9% / +7.9%
update(...) to 150 / 5000 before decode, restored after	+6.2% / −1.4%	+4.9% / +7.9%
scoped_update(...) with 50 / 5000 around decode	+6.2% / +5.9%	+4.1% / +6.5%

• Reading the limits through config::get on every needs_commit costs nothing measurable: this PR with defaults is within ±0.4% of its base.
• scoped_update and update give the same result, and peak memory during a 32k prefill is unchanged in every run, since the limits are back to their defaults before prefill.
• One small thing in the description: scoped_update("MLX_ENABLE_TF32"=0) isn't valid Python. The binding takes keywords, so it's scoped_update(MLX_ENABLE_TF32=0).

Thanks for this, it covers what #4562 was for.

Co-authored with Claude Opus 5.5

[zcbenz]

Thanks for verifying the change and for catching the python error!