CuTeDSL: nested comprehension breaks staged control flow later in the outer comprehension (NameError on the generator variable)

Description

_visit_Comprehension in cutlass/base_dsl/ast_preprocessor.py (~lines 2424-2428) sets session_data.generator_targets for the comprehension being visited, then unconditionally resets it to [] afterwards:

self.session_data.generator_targets = targets
ele_visitor(node)
self.session_data.generator_targets = []

When the element expression contains a nested comprehension, the inner _visit_Comprehension overwrites generator_targets with the inner targets and resets it to empty on exit. The outer comprehension's remaining element processing then runs with no generator targets, so any later construct that synthesizes staged control flow (e.g. an IfExp) treats the outer generator variable as a free/global name and fails.

Minimal repro (verified against nvidia-cutlass-dsl==4.7.0; the control version without the inner comprehension compiles and runs):

import cutlass
import cutlass.cute as cute

@cute.jit
def f(flag: cutlass.Boolean) -> cutlass.Int32:
    vals = [(len([0 for _ in range(2)]), (y if flag else -y)) for y in range(4)]
    return cutlass.Int32(sum(v[0] * v[1] for v in vals))

print(f(cutlass.Boolean(True)))

Actual error:

Caused exception: name 'y' is not defined
suggestion: Variables used inside staged control flow (for/if/while) must be defined before the
control flow region. Give the variable an initial value before the loop or branch.

The suggestion is misleading: y is the comprehension variable, not a user mistake.

Suggested fix

Save the previous value of session_data.generator_targets before visiting and restore it after ele_visitor(node) returns (a try/finally, which also protects the session state on exception paths), instead of resetting to [].

Start in cutlass/base_dsl/ast_preprocessor.py around _visit_Comprehension (approximately lines 2424-2428) and inspect how session_data.generator_targets is changed around ele_visitor(node). Run the minimal reproduction from the issue, then verify that nested comprehension processing restores the outer targets, including when element processing raises an exception.