[Refactor][Engram] Move the decode compile boundary to the op layer

[lcy-seso]

Problems

• EngramDecodeKernel registers its own torch.library custom op tileops::engram_decode with a hand-written fake, the last kernel-level custom op under src/tileops/kernels/; the compile boundary belongs to the op layer (ops-design.md § Compile Dispatch Boundary).
• EngramDecodeFwdOp already declares compile_boundary = True and is in the compile-contract registry, so the kernel only runs inside the op's generated operator and the nested custom op is an extra dispatcher round trip.
• The other kernel-level custom ops lost their decorators earlier but kept their bodies: 34 undecorated def _ fake functions, nine uncalled launch wrappers and fakes in DeltaNet and GLA, and comments claiming a custom op or torch.compile wrapper still exists.

Changes

• EngramDecodeKernel.forward allocates its outputs and launches the projection and step programs directly, with the same arguments, buffers and order as the deleted custom op body.
• Delete the undecorated fake functions, the uncalled wrappers, the imports only they used, and the stale custom-op comments across src/tileops/kernels/.
• Surface change: the operator torch.ops.tileops.engram_decode no longer exists; nothing in the repo called it, and the op's generated operator is unchanged.
• Eager outputs of EngramDecodeFwdOp are bit-identical (torch.equal) before and after on the three manifest rows and two extra rows (history shorter than the cache, unaligned d_mem).
• The kernel's launched programs and configs are unchanged, so no benchmark numbers; the nightly run covers it.
• Test node delta: 0 (no test file touched).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.