feat: support wandb project env

[firefighter-eric]

This PR allows the W&B logger to respect the WANDB_PROJECT
environment variable.

Previously, WandbLogger always used the explicit
project_name argument passed to wandb.init(), which
defaulted to DiffSynth-Studio. This meant that setting
WANDB_PROJECT in the environment did not change the W&B
project name.

With this change, WandbLogger reads WANDB_PROJECT before
calling wandb.init(), while preserving the existing default
behavior when the environment variable is not set.

Changes:

• Read WANDB_PROJECT in diffsynth/diffusion/logger.py
• Keep the default project name as DiffSynth-Studio
• No changes to run name behavior or other logger behavior

Validation:

• Ran python3 -m py_compile diffsynth/diffusion/logger.py

这个 PR 让 W&B logger 支持读取 WANDB_PROJECT 环境变量。

之前 WandbLogger 会把显式传入的 project_name 传给
wandb.init()，默认值是 DiffSynth-Studio。因此即使用户在环境
变量里设置了 WANDB_PROJECT，W&B project 名称也不会被覆盖。

这个改动会在调用 wandb.init() 前读取 WANDB_PROJECT。如果没
有设置该环境变量，则保持原来的默认行为不变。

变更内容：

• 在 diffsynth/diffusion/logger.py 中读取 WANDB_PROJECT
• 默认 project 仍然保持为 DiffSynth-Studio
• 不改变 run name 或其他 logger 行为

验证：

• 已运行 python3 -m py_compile diffsynth/diffusion/
  logger.py

[gemini-code-assist]

Code Review

This pull request updates the WandbLogger initialization in diffsynth/diffusion/logger.py to allow setting the Weights & Biases project name via the WANDB_PROJECT environment variable. The reviewer noted that this environment variable would override an explicitly passed project_name argument, and suggested only falling back to the environment variable if the default project name is used.

[Artiprocher]

@firefighter-eric Thanks for your contribution. We will add this parameter to the parsers immediately, and also support configure this parameter in SwanLab.

[firefighter-eric]

@firefighter-eric Thanks for your contribution. We will add this parameter to the parsers immediately, and also support configure this parameter in SwanLab.

Thanks for merging this and for planning parser / SwanLab support.

One related follow-up request: it would be very helpful if DiffSynth could log more training runtime metrics to W&B / SwanLab, not only the basic training metrics.

For example:

• step time / iteration time
• data loading time if available
• forward / backward / optimizer step time if available
• checkpoint saving time
• checkpoint path / checkpoint size
• samples per second or images per second
• GPU memory allocated / reserved / peak memory
• global step, epoch, learning rate, loss, gradient norm

These metrics would make it much easier to monitor training performance and diagnose runtime bottlenecks.