Add tail-adaptive f divergence to variational inference

[OleksiiBevza]

Description

Extension of variational inference with support for multiple divergence objectives.

Estimating and optimizing α-divergences require to use importance sampling, which may have large or infinite variance due to heavy tails of importance weights. In order to avoid this, researchers propose tail-adaptive f divergence that adaptively change the convex function with the tail distribution of the importance weights, in a way that theoretically guarantees finite moments,
while simultaneously achieving mass-covering properties.

Concretely, this PR adds support for:

• the standard KL objective
• the Rényi-α objective
• the tail-adaptive f-divergence objective

In particular:

• shared objective logic is added to the Gaussian VI code
• MFVI and FRVI now accept an objective argument
• the existing KL behavior is preserved as the default
• tests are updated for KL, Rényi-α, tail-adaptive f-divergence, and STL compatibility

This gives users a unified way to experiment with different VI objectives and compare behavior across scenarios, including standard KL-based optimization, α-divergence objectives, and more robust mass-covering tail-adaptive alternatives.

Based on:
(i) https://arxiv.org/abs/1810.11943
(ii) https://seunghan96.github.io/bnn/5.Variational-Inference-with-Tail-adaptive-f-Divergence/
(iii) https://github.com/dilinwang820/adaptive-f-divergence

[OleksiiBevza]

I'm also considering contributing implementations of two VI frameworks:
(i) Wasserstein Variational Inference: https://arxiv.org/abs/1805.11284
(ii) Operator Variational Inference: https://arxiv.org/abs/1610.09033

Both methods are theoretically motivated to be more robust and capable of better posterior approximations than standard VI, particularly for complex distributions. Would love to hear any design suggestions to keep the implementation clean :)

In the meantime will add an example with existing VI to sampling book.