pgw#1648: hardcut streaming fills to tensorfs

[PaulFidika]

Summary

• replace PGW streaming reads, pinned ring, copy walk, checkpoint image pread, warm torch copy, and cold varena refill submission with primitive destination/source maps handed to tensorfs
• require the native tensorfs store; delete the bridge fallback and streaming staging.py in the same cut
• preserve the construction census predicate/fence and admit the binding-carried expected census through the exact hook coordinated with th#2290
• keep the seam torch-free: strings, integer pointers/capacities/offsets, shapes, element width, layout token, and immutable file-range/hole records only

Current proof at 26b2e80

• real CPU native-fill pipeline is byte-equal across more than 100 tensors
• repeated fresh load after interpreter/config warmup: /proc/self/io write_bytes == 0
• 91 combined checkpoint, seam, streaming, lane-dtype, quantized, tied-weight, and census tests green
• warm B-to-A-to-B swap arm uses identical reservation addresses and records zero compile calls
• cold swap arm uses the same tensorfs client with complete file/hole records and honest destination byte accounting
• ruff and strict mypy green on touched source
• deletion census: no BridgeWeightStore, StagingPool, cudaMemcpyAsync, legacy _Placement/_walk, checkpoint _pread, torch warm copy, or varena refill submit remains in the cut paths; streaming/staging.py is deleted

Executed red/green evidence

• tensorfs planted mutation: forcing every requested arrangement to torch.contiguous@1 made the independent channels-last oracle fail at byte index 1; restoring the requested arrangement returned it green
• existing null/short destination, wrong-rank, unratified-layout, unknown-name, invalid-ledger, and partial-switch refusal arms remain green

Explicit blockers before ready/merge

• tensorfs PR changelog: gw#456 entry for 0.13.5 #153 merged as 1828c90 and v0.1.0 is tagged there; release run 32538259817 proved all gates, six wheel builds/smokes and two sdist installs green, then PyPI refused with invalid-publisher because no trusted publisher matches repo cozy-creator/tensorfs, workflow .github/workflows/publish.yaml, environment pypi. Both project and exact-version endpoints remain 404. Configure the PyPI publisher, then rerun failed jobs unchanged; do not retag/refire.
• tcg#63: torchcg's unversioned tensorfs Git source conflicts with PGW's exact 1828c90 draft source pin. Once 0.1.0 is public, torchcg must depend on exact tensorfs 0.1.0 and delete its source override; PGW will not weaken its pin.
• one real endpoint same-card/same-tree parity measurement remains required; no endpoint will be spent/refired blindly

Fence

• pgw#1645 PR pgw#1645 (3a/4): a worker on a lower layout RUNG reads as EARNING — and the decline that stands today names what is missing rather than dropping the wish #1190 merged before this branch rebased to dd39d02; its four files remain untouched
• th#2290 owns bind-contract decode/caller/protocol paths and is stacking on the narrow expected_census hook in this PR; it does not edit the streaming/checkpoint-juggle paths owned here

Tracker: pgw#1648