Report SKUAware storage reservation in planner stats

[hammad45]

Summary:
_compute_storage recognises FixedPercentageStorageReservation,
HeuristicalStorageReservation and InferenceStorageReservation, and falls
through to reserved_hbm_percent = 0.0 for anything else.
SKUAwareStorageReservation was not imported into stats.py at all, so every
successful SKUAware plan reported a reservation of zero.

Three things were wrong on each such plan:

• Reserved Memory: HBM: 0.0 GB / 0% and Planning Memory: 100%, regardless of
  what was actually reserved.
• The Dense Storage (per rank) and KJT Storage (per rank) blocks were
  suppressed entirely, because the same isinstance gate guards their rendering.
• Per-rank utilization was computed against (1 - 0.0) * hbm, overstating the
  headroom the planner really had.

Observed on a dry run of an APS model: the HeuristicalStorageReservation(25%)
arm reported 23.751 GB / 25% while the SKUAware arm of the same model reported
0.0 GB / 0%.

SKUAware reserves absolute bytes anchored to a fixed home SKU, so there is no
percentage to read off it. This reports the STATIC base --
model_base_bytes when set, otherwise margin_bytes, plus
runtime_overhead_bytes -- expressed as a fraction of per-rank HBM, mirroring
the existing FixedAbsoluteStorageReservation branch. Dense and kjt are
deliberately excluded: they render as their own blocks, and including them here
would count them twice.

Differential Revision: D116706171

[meta-codesync]

@hammad45 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D116706171.