Add cache and offload support for TracInAttributor

[brightXian]

Description

1. Motivation and Context

Implements cache + offload for TracInAttributor as discussed in #255.

2. Summary of the change

Added offload parameter with three modes: "none" (in-memory), "cpu", "disk". Reuses existing offload manager from block_projected_if.

attributor = TracInAttributor(..., offload="none")  # or "cpu" / "disk"
attributor.cache(train_loader)
score = attributor.attribute(test_loader)

Also changed attribute(train, test) → attribute(test, train=None) to support the cached path. All existing callers updated.

Benchmark

Here is a simple benchmark on MNIST LR, 10k train / 2k test, batch 100, 2 checkpoints, CUDA, projector dim 512. All three cache modes match no-cache baseline.

Mode	Avg (s)	Peak (MB)
No cache	61.81	829.93
Cache (none)	3.37	846.86
Cache (cpu)	5.93	787.52
Cache (disk)	6.05	787.52

The speedup comes from avoiding redundant train gradient recomputation. "none" is fastest by concatenating all cached grads for a single matmul; "cpu"/"disk" trade speed for lower memory by retrieving batch-by-batch.

3. Tests

• Unit test: cache + offload tests for all three modes in test_tracin and test_tracin_proj

[TheaperDeng]

LGTM