fix: improve quantization and runtime correctness

[ZX-ModelCloud]

Summary:

• Release Hessian partials.
• Safe sharded-load capability defaults.
• Instance loader/lm_head.
• ERNIE projection paths.
• GiB formatting.

Tests:

• Targeted: 92 passed, 35 skipped.
• Module-tree: 106 passed, 2 deselected.

[ZX-ModelCloud]

Reviewed commit 39971cd3bd214fc60f7a4a87a72eec6baf35589f.

Result: no blocking findings.

The changes are internally consistent:

• The fallback path resolves its compute device before releasing per-device Hessian partials, preserves the prior fallback arithmetic/results, and free() now drops lingering partial references.
• The nested getattr in kernel selection avoids eager evaluation of a missing legacy capability attribute while preserving SUPPORTS_SHARDED_LOAD precedence.
• get_model_with_quantize() now correctly honors definition-subclass loader and lm_head overrides through the instance.
• The ERNIE expert projection correction matches the surrounding gate_proj/up_proj/down_proj layout.
• math.prod keeps the column count as a plain Python int.

Validation performed in an isolated checkout:

• Changed/related test files: 60 passed, 31 skipped
• tests/test_looper_concurrency.py: 19 passed
• Full tests/module_tree: 106 passed, plus 2 failures that reproduce unchanged on origin/main (one existing None vs [] assertion and one local CUDA-initialization/environment failure)
• git diff --check origin/main...HEAD: passed
• Current GitHub CodeQL and Ruff checks: green

Non-blocking cleanup nit: in get_max_memory(), the two bare expressions reading active_bytes.all.peak have discarded results. Since peak values are not used, those expressions can be deleted rather than retained as no-ops.