I was affected by the 1kHz callback noise, and I found a solution that works great, at least for my application, so I wanted to share it here (as this thread helped me a lot to understand what was going on).
I designed a custom board based on the Daisy Seed, similar to the Terrarium, as in it has analog input and output buffers. I use it to run a NAM A2 model, an impulse response for cabinet simulation, and a reverb. The 1kHz whine was really annoying, no matter how I tried to ensure there were no ground loops, digital and analog ground were connected with a single low-impedance path, and the analog input power was stabilized. The load stabilization suggested by
@Newcastle in the previous comment (and in the DaisySeed docs too) was unfortunately not helping in my case. All the computation I am doing in the audio callback causes spikes in the power consumption once per callback (so, 1kHz), clearly visible with an oscilloscope on the Daisy Seed VIN pin, and unfortunately propagating to the analog part through the input buffer and possibly through the 4.5V audio reference.
Reducing the block size from 48 down to 8 or less would have sent the frequency in a range that is mostly filtered by guitar cabinets, and going down to 2 would go outside of the audible range, but that is not feasible in my application (or, at least, I could not make the callback efficient enough to allow for it).
What worked for me was digitally filtering the noise away. Here's the thing, this noise has three very important properties, at least in my case:
1. It mostly affects the circuit through its analog input
2. It aligns perfectly with the audio block (by definition, because it is caused by the audio callback itself)
3. It is constant across blocks: as long as I don't change the analog input gain, the noise always has the same pattern on every block.
This provides an opportunity to filter it: I keep 48 variables, each holding the running average of each one of the 48 samples in the block (across different invocations of the audio callback, trailing off over around 1 second). Then, I subtract such average from the input block. Effectively, this removes from the input signal any component that perfectly aligns with the block and is stable over time.
The noise is completely gone, one can only hear a small transient noise over 1 second when I tweak the analog input gain, that can be changed with a potentiometer in my board, but that's fine as this is not changed while playing. The tone is unaffected. Theoretically, an input signal very close to 1kHz or its multiples and relatively constant over time could be affected, but in practice guitar has too much transient and variation in phase/amplitude/pitch that I do not detect any negative impact, even if I play a B at the 19th fret (at around 988Hz it's the closest note to 1kHz), or even if I try to bend it to 1kHz. If anything was audible, a longer moving average would make it even less noticeable, at the (acceptable to me) cost of a slightly longer transient when changing input gain.
I know it might not be a viable option for all applications (for example, if your computation can change a lot while playing), and that fixing it in hardware would be better. If I design a new revision of the board, I might try using a stabilized 5V as audio reference instead of dividing 9V into 4.5V, and to isolate the analog power section even more. That said, this solution did save me from having to design an entirely new board, and for my application it has basically no practical downsides.
I hope it may help others too.