I was hoping to find some tool that already exists.
The short answer is, none exist yet. I see a number of hurdles to overcome first.1
For example, no one has yet characterized saturation for transcript discovery, i.e., the reads per cell needed to maximize transcript information recovery. This is a typical stepping-stone as a sequencing approach matures, and until it's done, it makes it unlikely that most scRNA-seq datasets are going to be appropriate vis-à-vis statistical power, even if there was a tool. That is, if people are only aiming for cell-type discrimination via gene-level markers, datasets will likely be underpowered to measure transcript level differences.
Another issue is the lack of any transcript-level scRNA-seq simulators. All the simulators out there that I know of produce gene-level matrices. Having a transcript-level simulator is crucial for benchmarking and characterizing the performance of differential expression calling methodologies. Without anything to test against, that leaves a huge blindspot in terms of estimating the uncertainty (false positive rates, sensitivity) that any tool might have. However, since some scRNA-seq models (e.g., Velten, et al. 2015, or Lopez, et al., 2018) are generative, perhaps this is shouldn't be too difficult to adapt to generating FASTQs.
Simulators are fine and all, but ground truth is probably the key hurdle of the enterprise right now. While there are definitely groups using smFISH to distinguish isoforms quantitatively on a single-cell level, there doesn't appear to be enough smFISH data out there to measure a transcriptome-wide scRNA-seq method against. Hopefully this will change in the coming year or two, as results from more highly-multiplexed single-molecule technologies become available.
1 This is just my opinion formed as I try to keep up with the literature, so I could be missing things. Comments/corrections certainly welcome!