Switch to a more modern microcontroller with better ADC and hardware-based timing (the micros() function with interrupts will give you less precise results). Maybe buffer data for a whole run and dump it at the end. This might be optional. At 115200, dumping 60 characters out will take nearly a quarter of your window.
The ADC on the 328P is actually quite good despite the resolution. Which we both know we can oversample it pretty easily for that resolution. Generally I have found that even over sampling the 328P ADC it is faster than an I2C external, not as fast as an SPI, but hey... As to the linearity of the Renesas ra4m1, I'm not sure. I'll probably have to do a linearity test myself to know since Google is failing me.
Another thing to think about in this build. Where to measure voltage for the input. I'm actually thinking that a 4 wire setup be used and measure at the ESC as close to the internal contacts as you can. This would then eliminate the wire length as a variable, and can be achieved with a rail to rail op amp and a handful of precision resistors. There is not a dyno that I know of that uses a 4 wire measurement, so maybe it is overboard...