Is Absolute/ Immediate Finality a goal in blockchain consensus or does some areas work better under probabilistic finality such as Ethereum's Pow and to be PoS? (I know that some networks using BFT mechanisms obtain Absolute/ Immediate finality. Is that not the best way to do it?)
"Best" depends on the goals.
Blockchain provides "finality" through a data structure and a process that makes it possible to reason about finality. That is a deliberately vague description because there are many variations on the structure, the process and the idea of confirmation.
BFT uses validators. They are privileged and they could (potentially) exert outsized influence, e.g. censorship, if they wanted to, so everyone has to trust that they won't abuse the privilege. In other words, the fast confirmations come at the cost of re-introducing a certain amount of centralization.
Since everyone needs to trust the validators, there is accountability and that implies that the operators will be identified - roughly the opposite of a network of distrusting, anonymous and potentially hostile nodes.
Proof-of-Work systems like Bitcoin and Ethereum came first, and they show that the entire thing is achievable without trust, accountability or privilege. It does so with comparatively less throughput/efficiency. In other words, trustless comes at the cost of performance.
It can be cognitively useful to consider two extremes:
Trustless and slow
Trustful and fast
What is "Best" actually depends on the application and the assurances the developer wants to be able to offer to participants. As a blockchain app developer, even a superficial understanding of requirements will usually make clear what sort of platform will be best. The platform will define many characteristics of any app before a line of code is written.
There is, of course, a lot of experimentation at the level of consensus and finalization. Ethereum itself, for example, innovates on proof-of-work with special treatment of ommar blocks. That makes it possible to reduce block time from 10 minutes to about 15 seconds without loss of efficiency. Still trustless, but faster than before.
Many variations on consensus attempt to disperse the centralized trust in validators present in algos like BFT. They tend to be on the fast side and trustful, but not quite as trustful as before.
Hope it helps.