Subgraph nodes with duplicate inputs silently drop connected values, causing cryptic missing argument errors

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

## Bug: INPUT_IS_LIST nodes fail with missing input argument after loading workflow from video metadata + browser refresh

### ComfyUI version
0.17.2 (commit 4a8cf359)

### Description
When a workflow is loaded from video-embedded metadata and the browser is then refreshed before running, nodes with `INPUT_IS_LIST = True` (specifically `ImageRebatch` / `RebatchImages`) fail with:

```
TypeError: ImageRebatch.rebatch() missing 1 required positional argument: 'images'
```

The input wire is visually connected in the graph. Running the workflow immediately after loading (without refreshing) works correctly.

### Steps to reproduce
1. Load a workflow from video metadata (drag video onto canvas or use Load Image node)
2. Refresh the browser
3. Run the workflow
4. Any `INPUT_IS_LIST` node with connected inputs will fail with a missing argument error

### Expected behavior
Workflow runs correctly after refresh, same as running without refresh.

### Actual behavior
`input_data_all` is missing the input key for `INPUT_IS_LIST` nodes, causing the call to `f(**inputs)` in `_async_map_node_over_list` to fail.

### Root cause hypothesis
The CacheProvider API introduced in `4a8cf359` (re-adding `af7b4a92` after revert) appears to have changed how cached node outputs are resolved and reconstructed for downstream nodes. After a refresh, stale cache entries are partially used to reconstruct `input_data_all`, and list-type inputs are dropped during that reconstruction for `INPUT_IS_LIST` nodes.

This did not occur in previous versions prior to the CacheProvider changes.

### Workaround
Run the workflow immediately after loading from video without refreshing, or manually dirty an upstream node (e.g. change a value and revert it) to force cache invalidation before running.

### Environment
- OS: Linux (Debian)
- GPU: RTX 4070
- Python: 3.13

### Actual Behavior

.

### Steps to Reproduce

.

### Debug Logs

```powershell
.
```

### Other

_No response_

Start by reproducing the refresh workflow with ImageRebatch/RebatchImages and inspect CacheProvider handling of input_data_all around _async_map_node_over_list. Trace why connected list-type inputs are dropped after cache reconstruction; done means the workflow runs after refresh without missing-argument errors and the regression is covered by an appropriate test.

Written by the indexing model from the issue text.