Pass temperature to draft sampler in dflash_generate

[shaun0927]

Closes #74.

Problem

In dflash_generate, the draft sampler is invoked without the
user-supplied temperature:

dflash/model.py:121 (draft):

block_output_ids[:, 1:] = sample(draft_logits)            # default temperature=0.0

dflash/model.py:134 (target):

posterior = sample(output.logits, temperature)

For any temperature > 0 the draft is therefore deterministic (greedy
argmax) while the target samples stochastically. Acceptance is decided
by a token-equality check
(block_output_ids[:, 1:] == posterior[:, :-1]), so the mismatch
artificially depresses acceptance and the accepted-token distribution
does not match the target distribution.

Reproduction without a model (verbatim copy of sample()):

import torch
def sample(logits, temperature=0.0):
    if temperature < 1e-5:
        return torch.argmax(logits, dim=-1)
    bsz, seq_len, vocab_size = logits.shape
    logits = logits.view(-1, vocab_size) / temperature
    probs = torch.softmax(logits, dim=-1)
    return torch.multinomial(probs, num_samples=1).view(bsz, seq_len)

torch.manual_seed(0)
logits = torch.tensor([[[2.0, 1.5, 1.0, 0.5]]])
draft  = sum(int(sample(logits).item() == 0) for _ in range(4000))
target = sum(int(sample(logits, 1.0).item() == 0) for _ in range(4000))
print(draft, target)   # 4000 1894

Fix

Pass temperature through to the draft sample():

block_output_ids[:, 1:] = sample(draft_logits, temperature)

This is the minimal change that puts draft and target on the same
sampling scheme. Acceptance is still token-equality (not
Leviathan-style rejection sampling), so dflash_generate still does not
provide an exact-distribution guarantee for temperature > 0; happy to
follow up with a docstring note or a proper rejection-sampling
implementation in a separate PR if useful.

Notes

• Greedy callers (temperature == 0, the default in the README
  examples) are unaffected: both branches go through torch.argmax.
• This is distinct from PR fix the sampling logic in model.py #67 which targeted the shape returned by
  sample() — that one would have broken downstream broadcasting; this
  PR touches only the call site.
• One of three small correctness fixes (also dflash_generate: final mask cleanup is value-based and drops legitimate output tokens equal to mask_token_id #73, MLX: _GDNStateCapture.rollback assumes "non-trimmable cache count == GDN layer count" without an assertion #75); each on its own
  branch so they can be merged independently.