fix(model): add explicit pretrain toggle to HIST initialization

[Ayush10]

Summary

• Add a pretrain parameter (default True) to the HIST model's __init__ to allow users to skip the pre-training weight-loading phase
• When pretrain=False, the weight-loading block in fit() is skipped entirely, allowing training from scratch
• Fully backward compatible — existing configs work without changes

Problem

The HIST model always executes the pretrained weight-loading logic during fit(), even when model_path is None. This creates unnecessary overhead (instantiating a base LSTMModel/GRUModel and transferring weights) and makes it impossible for users to train the architecture from scratch without the pretrain initialization step.

Changes

• qlib/contrib/model/pytorch_hist.py: Added pretrain=True parameter to __init__, stored as self.pretrain, and wrapped the weight-loading block in fit() with if self.pretrain:

Usage

```yaml

Existing behavior (unchanged) — pretrain weights are loaded

model:
class: HIST
kwargs:
pretrain: True # default, can be omitted
model_path: "path/to/pretrained_model.pkl"

New: skip pretrain entirely, train from scratch

model:
class: HIST
kwargs:
pretrain: False
```

Closes #2074