Several papers ([1], [2]) describe an optimised implementation of 2d convolution in which convolution operations are performed in the coefficient encoding space after which the output is converted to slot encoding.
Is there a “straightforward” way to use the routines inside of bootstrap to achieve this (as mentioned here and here)?
Theoretically they can certainly be used. Practically speaking, it is better to enable coefficient encoding for CKKS (so at least fresh ciphertexts could be generated using much cheaper encoding in the clear). I can certainly add an issue for OpenFHE v1.3.0 to add this functionality. You could your own encoding (or work with CoefPacked encoding) but this would require going deep in the internals of OpenFHE.