[InternalMetrics][Fix] fix dummy_fwd && refactor code && add CI

[nil0x9]

This PR involves several improvements for internal metrics:

1. (critical) Fixes error incurs when turning on internal metrics monitor and chunk loss is enabled. This is bc this PR adapted chunk loss to use torch.autograd.grad instead of torch.func.grad_and_value, which requires loss calculation to be performed on require-grad tensors. The original dummy forward in internal metrics monitor would perform loss calculation in no_grad mode, causing a runtime error. This PR removes unnecessary loss_ctx to avoid this.
2. Refactor: remove global vars from original implementation for that we found the recompile count is actually on par when these vars are moved to class attributes.
3. Add UT of internal metrics to catch aforementioned issue goes unnoticed.