[KDA] Eagerly load Paddle training imports

[huangjiyi]

Summary

Move the Paddle KimiDeltaAttention training-path imports out of ShortConvolution and CP convolution runtime methods so a single import fla loads the complete training closure before Paddle compatibility is disabled. The CP convolution PyLayer now returns gradients only for the Tensor inputs tracked by Paddle.

The regression test monitors __import__ and sys.modules across dense, padding-mask, varlen, and CP convolution forward/backward. It fails if FLA executes an import after bootstrap or adds/removes an FLA module key.

Test plan

• python -m pytest --confcutdir=tests/paddle -q tests/paddle/test_kimi_delta_attention.py: 6 passed
• python -m pytest --confcutdir=tests/paddle -q tests/paddle: 11 passed; the two existing dense dt_bias checks remain at error ratio 0.008083792 against tolerance 0.008
• Two-rank Paddle KDA CP: passed
• Training import monitor: 0 runtime imports from FLA and no FLA sys.modules changes after a 50-module bootstrap
• Static AST audit of 14 training-contract files: 0 nested imports
• python -m pre_commit run --files ...: passed

Benchmark / NCU

Not applicable; no kernel or numerical algorithm changes.

Breaking changes

None.