Extend mapped-servable staging beyond Q4_K/Q6_K (Q8_0, Q4_0/Q5_x, Q5_K, BITNET_B1_58)

[michalharakal]

Follow-up to #1189/#1190. StorageCapabilities.MAPPED_SERVABLE_DEFAULT is honest but small: dense F32, Q4_K, Q6_K. Every other encoding under WeightResidency.MAPPED heap-stages, and the plan now (correctly) charges it against the heap budget.

Per format the recipe is mechanical (proven by #1190): row-major C variant reading (o·blocksPerRow + b)·bytesPerBlock + JNI direct-buffer shim + BufferPackedTensorData scratch-decoder arm + entry in the servable set + parity test vs the feed-order kernel on permuted bytes.

• Q8_0 / Q4_0 / Q5_0 / Q5_1 / Q5_K — direct application of the recipe.
• BITNET_B1_58 (I2_S) — different: the load repacks stock GROUP layouts to sequential ([ternary-f32] Phase 4: GGUF I2_S (type 36) import with group→sequential repack, keep-packed BITNET_B1_58 #1140), and a repacked copy cannot page from the source file. Needs the repack cached once to an app-files sidecar and mapped back on subsequent loads (the NeoGPU-converter SEQUENTIAL case could map directly, minus the scale trailer question). The ternary LUT kernel is already row-major by construction, so no new C is needed — only the direct-buffer entry and the cache file.
• Keep AllocationResolver.explain / the servable set in lockstep with what loaders actually do — MappedBudgetPlanTest.encodings_nobody_serves_from_a_mapping_stay_heap_charged_even_under_MAPPED is the guard.

[fiwio-developer]

Priority evidence: any model with a non-256-multiple hidden/ffn dim carries k-quant-fallback Q8_0 tensors (SmolLM2-135M: most of its weight bytes), and under MAPPED those currently heap-stage in canonical order where — without prepack — only the decoding reference serves them: measured 48,771 ms/step vs 65 with a workaround (details in #1193). Q8_0 should be the first format this issue tackles, and per #1193's corollary the cheapest fix may be registering the row-major keys for heap storage too, which needs no new C at all for the mapped-servable formats and only the Q8_0 _rm kernel for the rest.