Fix nested hyperparameters clobbering user config during optimization

[musaaero]

Description

Replaces the shallow kwargs.update(sampled_hyperparams) in the Optuna objective()
with a recursive deep_update() that merges nested dicts key-by-key instead of
replacing them wholesale. Implemented as a local helper (the issue suggested pydantic's
deep_update; a local function avoids adding a dependency for one utility).
The helper deep-copies the base dict, which also fixes a latent leak: the previous
self._hyperparams.copy() was shallow, so nested dicts were shared across trials.

Motivation and Context

Fixes #431. The hyperparameter samplers return nested dicts (e.g.
{"policy_kwargs": {"net_arch": ..., "activation_fn": ...}}). The old dict.update()
replaced the entire user-specified policy_kwargs, silently dropping keys like
features_extractor_class / features_extractor_kwargs required for custom policies
(the reporter observed NaN actions as a result).

Note: read_hyperparameters() applies study hyperparameters from load_trial()
(--trial-id) via the same shallow update() pattern and has the same issue — happy
to send a follow-up PR for that path if this approach is approved.

Types of changes

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I've read the CONTRIBUTION guide
• I have updated the changelog accordingly
• I have updated the tests accordingly
• I have checked the codestyle using make check-codestyle and make lint (CI)
• I have ensured make pytest and make type both pass (CI)

AI assistance disclosure: I used AI assistance to draft parts of this change.
I have reviewed and understood every modified line and can explain the reasoning.