feat: refactor four attention converters with TRT add_attention_v2()

[zewenli98]

Description

This PR refactors four attention converters with TRT add_attention_v2(). Additionally, since newer torch rejects passing both is_causal=True and explicit attn_mask to SDPA, which aligns with TRT, some scenarios are gated in validators and the unreachable branches are pruned.

Fixes # (issue)

Type of change

• New feature (non-breaking change which adds functionality)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-17 18:09:43.975250+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-17 18:10:04.835218+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-17 18:09:44.412631+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/lowering/passes/eliminate_sym_min_int64_max.py	2026-07-17 18:10:05.940612+00:00
@@ -2,11 +2,10 @@

import torch
from torch.fx import GraphModule, Node

from .pass_utils import clean_up_graph_after_modifications
-

_INT64_MAX = 2**63 - 1
_SYM_MIN = getattr(torch, "sym_min", None)

[narendasan]

This is quite high? Will it be stable in practice?

[zewenli98]

The relaxation is mainly for TRT-RTX. As I talked with @SandSnip3r, TRT-RTX 1.6 is going to fix the accuracy issue but it's not released yet. So I skipped these tests for now and changed to rtol=1e-1, atol=3e-1 for the single slightly mismatched element on standard TRT:

Mismatched elements: 1 / 9723904 (0.0%)
Greatest absolute difference: 0.2071533203125 at index (0, 0, 80751) (up to 0.2 allowed)
Greatest relative difference: 2.931640625 at index (0, 0, 80751) (up to 0.1 allowed)

[narendasan]

Should this be a lowering pass?

[zewenli98]

I'm not sure if this is general enough to make it a lowering pass now because I only observed in the query of attention in some model implementations like HF Qwen. Maybe we can keep it in the attention converters now. If we observe more cases later, we can make it a lowering pass then.