[Priority] [Bug] removing hardcoded local file paths and standardizing them

[Varun-sai-500]

This PR modernizes configuration handling and data transformation utilities to improve compatibility, maintainability, and robustness across training and deployment pipelines.

Key Changes:

1. Path Handling & Configuration Refactor
   Introduced BASE_DIR and standardized absolute path construction across configs
   Replaced hardcoded relative paths with os.path.join + os.path.abspath
   Refactored dynamic imports using f-strings
   Consolidated directory structure initialization using a common ROOT
   Updated logging paths:
   DIR_TB_LOG now derives from DIR_LOG
   Simplified directory creation with exist_ok=True

2. Model Configuration Improvements
   Updated all MODEL_ENCODER_PRETRAIN paths to use absolute resolution
   Ensured consistent path handling across all model configs
   Reordered initialization where necessary to avoid dependency issues

3. Semantic Cleanup
   Replaced sentinel values:
   TRAIN_LONG_TERM_MEM_GAP = 9999 → None
   TEST_LONG_TERM_MEM_GAP = 9999 → None
   Improves clarity and avoids implicit magic-number behavior

Important:
This refactor actually solves a problem because normal running the file (without flags) and it will just take it from here, and these configs have those vos02, it's incredibly harder to debug, so standardized to dataset paths
One noticeable change is I have removed data_wd and VOS2 from path because it really doesn't belong the project repository, it's local, slipped into production.

[Varun-sai-500]

@z-x-yang Can you please review?

[z-x-yang]

Thanks — this PR is much smaller than #96 and the direction is reasonable, but I don’t think it is merge-ready yet because the new path resolution is currently incorrect.

Main blockers:

1. In configs/default.py, BASE_DIR is already <repo>/configs, but ROOT = os.path.abspath(os.path.join(BASE_DIR, '..', '..')) resolves to the parent directory of the repo, not the repo root. When I instantiate the config, DIR_ROOT / DIR_DATA point outside the checkout, so datasets, results, logs, and checkpoints would be written/read from the wrong place.

2. In the model configs, BASE_DIR is <repo>/configs/models, but paths like:
   os.path.join(BASE_DIR, '..', 'pretrain_models', ...)
   resolve to <repo>/configs/pretrain_models/..., not <repo>/pretrain_models/....

So the current refactor would break config-driven training/eval by looking for datasets and pretrained encoder weights in the wrong locations.

The None change for *_LONG_TERM_MEM_GAP looks directionally fine with the engine guard, but it should be tested after the path issues are fixed.

Suggested fix: define a single repo root helper once, e.g. from configs/default.py, and use the correct number of .. levels consistently. Then smoke-test one training config and one eval/demo config to print the resolved paths before merging.

[Varun-sai-500]

fixed those issues, must be good by now @z-x-yang