perf: batched NLI inference + shared model instances (closes #330)

[mohitcek]

Description

NLI scoring is the main latency bottleneck for Semantic Entropy, Semantic Density,
Noncontradiction, Entailment, and Number of Semantic Sets (#330). This PR speeds it
up in three ways:

1. Batched inference. NLI.predict_batch() scores many premise/hypothesis pairs
   per forward pass (configurable nli_batch_size, default 32) under
   torch.no_grad(). All callers now collect their pairs first and score them in
   bulk: clustering batches each round's comparisons, ConsistencyScorer and
   SemanticDensity batch across all prompts, and the longform claim/graph scorers
   batch their claim × response loops. predict(premise, hypothesis) remains as a
   single-pair wrapper for backward compatibility.

2. One model copy instead of 2–4. Default BlackBoxUQ() loaded the ~1.4 GB NLI
   model twice; its scorers now share a single injected instance.
   SampledLogprobsScorer no longer re-instantiates SemanticEntropy /
   SemanticDensity / SentenceTransformer on every evaluate() call.

3. Correct model inputs. Proper tokenizer(premise, hypothesis) pair encoding
   replaces the hand-concatenated " [SEP] " idiom, and real token-level truncation
   (512 tokens for the default model) replaces character slicing. Also fixed: the
   pair cache used collision-prone f"{r1}_{r2}" string keys (now tuples).

Extras: opt-in fp16 on CUDA/MPS, optional device_map="auto" via a new
uqlm[accelerate] extra, and ConsistencyScorer now honors device.

Benchmarks

Apple M-series, torch 2.10; before = main. Reproduce with python -m uqlm.nli.benchmark
(its sequential/pipeline modes also run on old releases). CUDA numbers welcome —
batching should gain more there.

Benchmark	Device	Before	After
BlackBoxUQ noncontradiction+entailment, 64 pairs	mps	15.5s	3.5s (4.4x)
BlackBoxUQ noncontradiction+entailment, 64 pairs	cpu	15.4s	4.9s (3.1x)
BlackBoxUQ 3 NLI scorers, 8 prompts × 5 samples	mps	8.6s	2.6s (3.3x)
Raw NLI throughput, short texts	cpu	8.4 pairs/s	33.0 pairs/s (3.9x)
NLI model copies, default BlackBoxUQ()	—	2 (~2.8 GB)	1 (~1.4 GB)

Behavior notes — scores change slightly

The old " [SEP] " concatenation leaked a stray whitespace token into the model
input (and corrupted inputs for non-DeBERTa nli_model_name values). The new pair
encoding matches deberta-large-mnli's MNLI training format exactly, so NLI scores
are not bit-identical to previous releases: noncontradiction typically moves by
~1e-4..1e-3; entailment can move more on QA-style text. Texts beyond 512 tokens
were previously fed to the model over-length; they are now truncated correctly.
Test fixtures were re-scored with the same stored responses.

Batched vs sequential inference is a numeric no-op (parity test included), and
batched clustering was verified pair-for-pair identical to the old sequential loop.

Additional bug fixes

Five verified bugs found while working in these code paths, each with a regression
test that fails on main:

1. Tuner optimized the wrong direction for log loss — a "logloss"/"log_loss"
   key mismatch made the ensemble tuner maximize log loss (uqlm/utils/tuner.py)
2. Functional entropy assigned rotated cluster probabilities — [j-1] index
   wraparound; token-probability-weighted code-UQ scores were wrong (uqlm/code/entropy.py)
3. UQEnsemble.score() ignored its logprobs_results argument — fresh instances
   crashed, reused instances silently scored stale logprobs; adds
   sampled_logprobs_results (uqlm/scorers/shortform/ensemble.py)
4. LongTextUQ(mode="matched_unit") always crashed — float() applied to the
   list returned by the cosine scorer (uqlm/longform/luq/matched_unit.py)
5. LLM-judge entailment retry path never worked — retried with DataFrame column
   names as prompts, then crashed indexing a list; case-sensitive yes/no parsing
   (uqlm/nli/entailment.py). Its test class was converted to
   IsolatedAsyncioTestCase — the async tests previously never executed.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update
• Dependency update

Checklist

• Tests added or updated for all changed behavior
• Docstrings updated for any new or modified public API
• Type annotations added for any new or modified functions
• ruff check and ruff format pass locally