docs: VMI/PTODSL feature-gap reproducers (scalar GM, vbrc, 1PT, UE8M0, Persistent)

[learning-chip]

Summary

Adds minimal VMI/PTODSL feature-gap reproducers under docs/feature-gaps/vmi/ for five PTOAS capabilities needed by high-throughput quantization and residual-mix vector kernels. Each gap directory is a documentation fixture (not a lit test) with the same structure:

• README.md — algorithm need, current failure, desired PTOAS behavior
• buggy_vmi.pto — compiling-today workaround IR
• desired_vmi.pto — idiomatic VMI using the missing feature
• target_mi.pto — desired VPTO / pto.mi shape after lowering

Gaps

1. Scalar GM dcache bypass (01_scalar_gm_dcache_bypass) — per-block scales need cheap scalar GM writes without bulk MTE.
2. Compact inverse vbrc (02_compact_inv_vbrc) — fan reciprocal scales without padding to a full chunk before dist_mode=brc.
3. Direct scalar reduce store (03_direct_scalar_reduce_store) — store one f32 reduce result without 8-slot padding / enable 1PT_B32.
4. Packed UE8M0 SF (04_packed_ue8m0_sf) — pack_factor=2 UE8M0 scale factors instead of unpacked ui8.
5. Persistent stages / block_k (05_persistent_stages_block_k) — multi-buffer Persistent stages and block_k > 512 for long-K tiles.

Test plan

• Review each README.md for accuracy vs current feature-vmi lowering
• Optionally feed desired_vmi.pto / buggy_vmi.pto through ptoas to confirm documented failure modes
• No runtime/lit expectations (docs-only)

[learning-chip]

Moved to hw-native-sys#997