…IDIA-NeMo#2221)

NeMo-RL pins ``transformers==5.3.0`` but currently declares
``accelerate>=0.26``. Versions of accelerate older than 1.1.0 are
incompatible with transformers 5.3.0's call sites that pass the
``_is_hf_initialized`` keyword to ``Parameter.__new__``, producing a
runtime crash:

    TypeError: Parameter.__new__() got an unexpected keyword
    argument '_is_hf_initialized'

The currently locked version (1.10.0) is unaffected because the
resolver was constrained from elsewhere, but a fresh resolve in a
different environment can pick a too-old accelerate and fail at
runtime. Tighten the floor so the constraint accurately reflects
transformers 5.3.0's real requirement.

uv.lock was regenerated with ``uv lock``. The only semantic change is
the matching ``accelerate`` specifier in nemo-rl's
``[package.metadata.requires-dist]`` block; the resolved version
(``accelerate==1.10.0``) and every other package version are
unchanged. The remaining lockfile churn is resolution-marker
normalization from a newer uv release and is benign.

Refs: NVIDIA-NeMo#2221
Signed-off-by: lonexreb <[email protected]>

The previous commit on this branch claimed `accelerate>=1.1.0` would
fix the ``_is_hf_initialized`` ``TypeError`` from NVIDIA-NeMo#2221. That was
wrong — verified by reading accelerate's source on the
huggingface/accelerate repo:

* v1.10.0 ``register_empty_parameter`` (src/accelerate/big_modeling.py:129)
  has no handling for the ``_is_hf_initialized`` kwarg, so transformers
  5.3.0's ``Parameter.__new__`` call still crashes.
* v1.13.0 ``register_empty_parameter`` (line 131) pops ``_is_hf_initialized``
  from kwargs and re-applies it after parameter creation.

So the actual floor required to unblock transformers 5.3.0 is
``>=1.13.0``. Bump pyproject.toml accordingly and regenerate uv.lock —
``accelerate`` moves from 1.10.0 to 1.13.0; no other versions change.

Refs: NVIDIA-NeMo#2221
Signed-off-by: lonexreb <[email protected]>