Fix slice_scatter open end

[jloftin-nv]

Description

Export encodes open-ended slice ends as 2**63-1. The converter feeds that into np.arange / broadcast_to without clamping to the dim size, causing “array is too big” or broadcast errors.

Clamp end (and related bounds) to the input length along dim before building NumPy indices, matching aten runtime behavior.

Type of change

Please delete options that are not relevant and/or add your own.

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• [ X] My code follows the style guidelines of this project (You can use the linters)
• [ X] I have performed a self-review of my own code
• [ X] I have commented my code, particularly in hard-to-understand areas and hacks
• [X ] I have made corresponding changes to the documentation
• [X ] I have added tests to verify my fix or my feature
• [ X] New and existing unit tests pass locally with my changes
• [ X] I have added the relevant labels to my PR in so that relevant reviewers are notified

[apbose]

Dynamic sliced dimensions remain broken. The clamp is skipped when dim_size == DYNAMIC_DIM The code then reaches np.arange(0, 2**63-1), causing compilation failure or a huge-allocation error. The existing dynamic test varies batch and other dimensions while keeping the sliced dimension fixed at 64, so it does not cover this. We should include this test

[jloftin-nv]

Should we land dynamic support here or keep in torch and follow up?

[apbose]

yeah we can have it another PR. Just that we should have a validator then

[apbose]

_needs_dim_size() misses a case when the dim_is_static=False . A positive literal bound slips through unchecked.

resolve_slice_scatter_write does correctly attempt to guard the dynamic-dim case:
if not dim_is_static and (_needs_dim_size(start) or _needs_dim_size(end)):
return None, None, None, KVWriteStatus.DYNAMIC_DIM_SIZE
but _needs_dim_size() only flags three of the four bound shapes that actually depend on the dim's unknown size:
repro: slice_scatter(cache, update, dim=2, start=0, end=40, step=1) where dim 2's declared dynamic range is 8 to 32

bound	_needs_dim_size(bound)	caught?
None	True	yes
negative (-4)	True	yes
open-end sentinel	True	yes
positive literal (40)	False	no

I think we should widen the _needs_dim_size so concrete bound on dim isnt considered safe