[API Compatibility] Change compatibility apis to ChangePrefixMatcher, inject paddle.enable_compat() -part

[Manfredss]

PR Docs

For the 23 torch.* APIs that are aligned through paddle.compat.*, switch the
converter to the "prefix-only + enable_compat" strategy instead of mapping each to
paddle.compat.X explicitly:

1. api_mapping.json — change those 23 entries from
   ChangeAPIMatcher → paddle.compat.X to {"Matcher": "ChangePrefixMatcher"}, so
   torch.X(...) converts to paddle.X(...) with arguments untouched
   (e.g. torch.sort(a, dim=-1) → paddle.sort(a, dim=-1)).
2. transformer/import_transformer.py — inject import paddle +
   paddle.enable_compat() right after the imports of any converted file, so the
   prefix-only paddle.X calls resolve to the torch-aligned paddle.compat.X
   implementations at runtime. (Included explicitly + deduped so it is valid even
   when torch was imported as a submodule alias, e.g. import torch.nn as nn.)
3. tests/apibase.py — paddle.enable_compat() flips global state (an
   import torch → Paddle proxy + paddle.* aliases). Disable it before each
   PyTorch reference run (_ensure_paddle_compat_disabled()), so the reference
   always executes against real PyTorch, including across multiple run() calls in
   one test. (No-op until Paddle is first imported, so it does not change torch/paddle
   import ordering.)

PR APIs

torch.allclose
torch.equal
torch.max
torch.median
torch.min
torch.nanmedian
torch.nn.AvgPool1d
torch.nn.AvgPool2d
torch.nn.AvgPool3d
torch.nn.Linear
torch.nn.MultiheadAttention
torch.nn.softmax
torch.nn.Unfold
torch.nn.functional.linear
torch.nn.functional.log_softmax
torch.nn.functional.pad
torch.nn.functional.scaled_dot_product_attention
torch.nn.functional.softmax
torch.nn.functional.unfold
torch.seed
torch.sort
torch.split
torch.unique

related Paddle PR:

• [API Compatibility] enhance paddle.enable_compat -part Paddle#79391

Co-authored by Claude.