Trust me, that’s all answered inside the documentation and posts you get when following the links I provided.
TL,DR: There is no separate LDR denoiser training network anymore.
Is there any advantage using the denoiser before the host application performs tone-mapping rather than after ?
That is how you need to do it when using when using HDR denoiser. Tonemapping must happen afterwards with that.
is the HDR denoiser simply the LDR denoiser surrounded by tone-mapping and un-tone-mapping (inverted tone-mapping)
Absolutely not. That wouldn’t work.
Does the HDR denoiser use internally HDR value in the denoising ML model ?
Yes, and the valid range for the HDR input values is [0, 10,000] That was always the case.