Fix schedule_from task cleanup before start

[fallintoplace]

Problem

schedule_from releases the owning pointer for its continuation task during connect(), but the operation state only keeps a raw pointer after that. If the operation is destroyed without calling start(), the task never reaches the queue and its device and pinned allocations are not freed.

Fix

• Keep the continuation task owned by the operation state until the inner operation is started.
• Release ownership immediately before calling start(inner_op_).
• Keep cleanup correct when setup reports a CUDA error or connecting the inner sender unwinds.

Test

• Added a regression test for destroying an unstarted operation.
• Added a setup-error regression that fails the environment allocation after storage and task allocation succeed, starts the operation, checks cudaErrorMemoryAllocation, and verifies balanced cleanup.
• clang-format-21 --dry-run --Werror
• CMake configure and build with CUDA disabled

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.