Show requested slots for terminating sessions in the admin session list

## Problem

A 120-node batch session in `TERMINATING` shows **8 GPU / 64 CPU / 1544 GiB** in the admin session list, while the same session showed 960 / 7680 / 185280 GiB while `RUNNING`.

The manager's `ComputeSessionNode.occupied_slots` is the sum of kernel allocations that have **not been freed yet** (`resource_allocations.free_at IS NULL`). When 119 kernels are already reclaimed and one node is stuck, the session-level value shrinks to a single node's worth. The WebUI (`SessionSlotCell.tsx`) renders that value as-is and only falls back to `requested_slots` when `occupied_slots` is empty, so the list looks like the session lost its resources.

Reported in the Teams "AI FM" thread (see web link).

## Fix

In `SessionSlotCell`, prefer `requested_slots` over `occupied_slots` while the session is in a transitional status (`TERMINATING`, and the creating-phase statuses `PREPARING` / `PREPARED` / `PULLING` / `CREATING` / `SCHEDULED` / `PENDING` / `RESTARTING`), so the AI accelerator / CPU / memory columns keep showing the session's full request (960 GPU) until it is actually `TERMINATED`. Progress of the reclaim itself is surfaced by the status badge (follow-up issue).

## Acceptance

- A multi-node `TERMINATING` session shows the same accelerator / CPU / memory numbers it showed while `RUNNING`.
- `RUNNING` / `TERMINATED` sessions are unchanged.
- `bash scripts/verify.sh` passes.

JIRA Issue: FR-3922

Start in SessionSlotCell.tsx and trace how slot values are selected for session statuses. Check the handling of TERMINATING and the listed creating-phase statuses against the unchanged RUNNING and TERMINATED cases. Confirm that multi-node terminating sessions retain their requested accelerator, CPU, and memory values, then run bash scripts/verify.sh.

Written by the indexing model from the issue text.