feat(models): make Inkling implementation standalone (#3358)

* fix(pp): preserve VLM media cursor with static metadata

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* fix(ci): read static PP metadata in checkpoint parity

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* fix(checkpoint): ignore base FP8 metadata on resume

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* ci(vlm): extend Mistral4 checkpoint timeout

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>

* feat(models): make Inkling implementation standalone

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(models): align Inkling attention dtypes

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(models): preserve Inkling residual dtype

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(models): make Inkling training pipeline-safe

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

---------

Signed-off-by: Yuhe Zhang <yuhez@nvidia.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>
Co-authored-by: Yuhe Zhang <yuhez@nvidia.com>