fix(dflash): add sglang 0.5.14 post-load init to the DFlash target model

[jessiewei7]

Summary

On current main (which pins sglang==0.5.14 since #605),
prepare_hidden_states.py --model-type dflash crashes immediately:

File ".../dflash_target_model.py", line 141, in _extend
    batch = ScheduleBatch.init_new(req_to_token_pool=self.model_runner.req_to_token_pool, ...)
File ".../sglang/srt/managers/schedule_batch.py", line 1866, in init_new
    device=req_to_token_pool.device,
AttributeError: 'NoneType' object has no attribute 'device'

Root cause

sglang 0.5.14 refactored ModelRunner: __init__ no longer allocates the
memory pool. Pool allocation moved to explicit post-load calls
(alloc_memory_pool() → init_memory_pool() in
model_runner_kv_cache_mixin.py), normally invoked by the
Scheduler/TpModelWorker. SpecForge drives the ModelRunner directly through
its standalone SGLangRunner, so those calls never happen and
req_to_token_pool / token_to_kv_pool_allocator stay None.

#605 fixed exactly this for eagle3_target_model.py (added the three
post-load calls with a detailed comment) but did not touch
dflash_target_model.py, leaving the DFlash extraction path broken on the
very sglang version the repo now pins.

Fix

Mirror #605's eagle3 fix in SGLangDFlashTargetModel.from_pretrained():

model_runner.alloc_memory_pool()
model_runner.init_attention_backends()
model_runner.init_cuda_graphs()

(init_cuda_graphs() captures nothing here because DFlash sets
disable_cuda_graph=True; it still builds the EagerRunner that forward()
needs.) 4-line diff, identical pattern to the already-merged eagle3 change.

Since this is the second target model needing these exact calls, hoisting
them into SGLangRunner itself would prevent a third recurrence — but that
changes the pattern #605 established, so this PR keeps the minimal call-site
fix and leaves that refactor as a possible follow-up.

Validation

Patched and ran prepare_hidden_states.py --model-type dflash on
GLM-5.2-FP8, multi-node tp8, sglang 0.5.14: the generation loop completed
(Processed 176 / Filtered 24 short) and hidden-state .ckpt.gz files were
written correctly. Without the patch, the run crashes at the first batch as
shown above.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[jessiewei7]

@FrankLeeeee PTAL — dflash training crashes on main with sglang 0.5.14 (same pattern as #605). Companion fixes: #663 #659 #665.