revert: restore pre-#925 trainer execution with API-name aliases

[bradhilton]

Agent: Tarkin

Problem: #925 sends every CUDA forward through versioned graphs, CPU snapshots and a placement ladder, with no way to turn it off. It cost throughput, caused the 062 MemoryError (#1096), and needed #1081, #1086 and #1099 to price itself. No current experiment uses the shared-service features it was built for.

Behavior after this change:

• Trainer execution: it's pre-feat: add versioned trainer graphs and logical-rank APIs #925 SPMD again. trainer_rank/ and megatron/ match feat: add versioned trainer graphs and logical-rank APIs #925's parent 652e0f085, apart from the independent fix(megatron): skip TE grouped GEMM when every group is empty #1089 and the fixes below. The five trainer commits between deb95f3f and that parent stay.
• Fixes re-applied on the restored path: fix(trainer-rank): preserve dense no-grad floors beside adapter backward peaks #1067, fix(trainer-rank): reserve eager head workspaces in admission and replay #1068, fix(trainer-rank): price TP2/TP4 sequence-parallel recompute term by term #1075, fix(trainer-rank): retain HybridEP planner replay dimensions #1088 and fix(trainer-rank): combine cross-entropy target and softmax gradients in FP32 #1105 (FP32 head gradients and their pricing). Adapted: fix(trainer): honor oversized admission for supported EP plans #1080 and fix(trainer): require coordinated EP oversized admission #1084, without the CPU-offload checks, and fix(trainer): remember reloadable immutable checkpoint sources #1077, which Caladan's frozen reload (#1501) calls.
• Independent fixes from inside feat: add versioned trainer graphs and logical-rank APIs #925, kept:
  • GPT-OSS export trims packed gate/up rows in their interleaved order (46c911011). Without it, padded packed exports kept padding and dropped live rows.
  • The native export builder validates packed gate/up shapes (63b91ca90, b745db5a9).
  • Distributed export runs local collection before any metadata collective, with separate error phases for packed and dense metadata (b745db5a9).
• Dropped: fix(trainer-rank): credit staged adapter gradients and shared captures in graph admission #1081, fix(trainer-rank): preserve graph pricing in oversized fallbacks #1086 and fix(trainer-rank): honor oversized model output placement #1099, which only price or place feat: add versioned trainer graphs and logical-rank APIs #925 graphs and outputs.
• Aliases: the v1 names are aliases. forward_batches, forward and reduce are the restored methods, and backward(loss) matches loss.backward(). options= raises TypeError on every method and the constructor, even as options=None.
• Removed: client-side training, logical-rank callbacks, heads, ForwardOptions, ModuleHandle, CPU offload and replay, staleness corrections, and the cloudpickle dependency.
• Tokenizer: Preserve literal think markers in thinking-off histories #967's thinking-off workaround is restored from 652e0f085, per Brad's decision. On current real paths it stays inactive, because Preserve literal reasoning tags in assistant history #1009's normalization already produces the template it gates on; three tests cover the legacy template directly. Otherwise tokenizer and chat-template behavior is unchanged from main.

Validation:

• Local checks (GPUs hidden, using the quality-checks commands):
  • prek (ruff with the locked 0.15.15, ruff-format, ty, uv-lock-check) passes.
  • Megatron lightweight: 1,604 passed, 14 skipped, with only the known memory test failing.
  • Tokenizer and trajectory suites: 1,966 passed.
  • Alias tests pass.
• Caladan: all 16 GPT-OSS export cases from caladan#1995 pass, including the 4 padded cases that failed on the previous head.
• GitHub CI: the required 2×H200 TrainerRank GPU validation passes. In quality-checks, the only failure is test_tensor_parallel_heads_fit_their_charge_without_a_spare_chunk, which fails the same way on main. It's a real head-checkpoint memory bug that fix(trainer-rank): keep head row normalization outside the statistics checkpoint #1120 is fixing, and CI skips the unit step after it.

Limitations:

• That memory bug is present here, as it is on main, until fix(trainer-rank): keep head row normalization outside the statistics checkpoint #1120 lands.
• There's no dedicated GPU qualification against deb95f3f yet. The comparison long runs started from this revert are the first one.
• Caladan adopts this in caladan#1995: per-rank trainer.queue lists, no options=, and ports of #1501, #1663, #1741 and #1745.
• Planner reports spooled under the feat: add versioned trainer graphs and logical-rank APIs #925 pin won't validate on this code, because of their operation names and memory_placement field.