fix issue #83, Fix ObjaverseXL compatibility & improve CLI usability for Training prepare

[yihanz0107]

1.Fixes #83.

2.Fix missing "output_dir" for download()
Logic added:
parser.add_argument('--output_dir', type=str, default=None,
help='Directory to save the metadata')
opt.output_dir = opt.output_dir or opt.root

3.Fix other compatibility issue in 'data_toolkit/asset_stats.py', by automatic judgment

Validation:
I have validated these changes by running the full end-to-end pipeline on ObjaverseXL sample data (download -> dump -> build_metadata -> dual_grid/voxelize -> encode), and confirmed that all steps now execute without errors.

(trellis2) yihan@EZ-ubuntu:~/projects/TRELLIS.2$ python data_toolkit/build_metadata.py ObjaverseXL --root datasets/ObjaverseXL_sketchfab
Loading previous metadata...
Statistics:

• Number of assets: 168307
• Number of assets downloaded: 1
• Number of assets with render conditions: 1
• Number of assets with mesh dumped: 1
• Number of assets with PBR dumped: 1
• Number of assets with asset stats: 1
• Number of assets with dual grid:
  • 256: 1
  • 512: 1
  • 1024: 1
• Number of assets with PBR voxelization:
  • 256: 1
  • 1024: 1
  • 512: 1
• Number of assets with sparse structure latents:
  • ss_enc_conv3d_16l8_fp16_64: 1
• Number of assets with shape latents:
  • shape_enc_next_dc_f16c32_fp16_512: 1
  • shape_enc_next_dc_f16c32_fp16_1024: 1
• Number of assets with PBR latents:
  • tex_enc_next_dc_f16c32_fp16_1024: 1
  • tex_enc_next_dc_f16c32_fp16_512: 1

[yihanz0107]

@microsoft-github-policy-service agree