This is a though question that is surfing at the limits of meta/political wars within the cryptography community.
SafeCurves is a good resource, but it's very opinionated about what "safe" means.
As an example
secp256k1, the curve used by Bitcoin and Ethereum to guarantee their security is deemed "unsafe" by the SafeCurves team, while it currently has over a 1000+ B$ "bug bounty" in the form of BTC and ETH that could be hacked if you could break it...
So, no, it's not because SafeCurves says something is not "safe" that it means it's "broken". NIST curves could be better, sure, but they still work and do their job and (as far as we are aware of) there are no effective, practical attacks against them that would threaten systems that are currently using them, as long as they are "properly" implemented...
Which means they aren't broken, and I wouldn't personally say they aren't "safe" for usage in 2021. But if I'm designing a new system where I can choose whatever key agreement mechanism or signature algorithm, they also aren't my first choice.
A huge problem when you're working in software (and hardware) development is "compliance", and the NIST are the ones issuing the "standards" that are considered as the holy grail in the compliance world, I'm afraid.
So, while they are currently working on new drafts (e.g. the draft for FIPS 186-5 is currently including Ed25519 as a signature algorithm) containing more "modern" curves, they still haven't yet updated their recommendations and standards and as such many entities are mandated to rely on their good old recommendations such as the P-256 curve.
Also, it's important to notice that GnuPG is following the OpenPGP specifications and is trying to be a widely interoperable, inter-compatible tool. As such it supports a LOT of different schemes, amongst which some are older than other, some are more or less used.
There is a draft including Ed25519 and X25519 support in PGP that is currently open and has been actively worked on for the past few years:
The same compatibility issues also arise with PGP: if you want to be able to easily communicate with others using it, it's best to use something others are also using/supporting.
So, PGP will get there at some point, but it's a slow process and as long as these old curves and standards aren't practically (or at risk of being) broken, this slow process likely won't accelerate.
Now, should you change your PGP keys if you're using P-256 or another curve whose "cost for rho" is above 2^100 in that list: https://safecurves.cr.yp.to/rho.html ?
I would say "No". You don't need to worry for now, at least for the next 5 years, possibly more. But Quantum Computers might change this story in a few more years shall they be as effective as we expect them to be at breaking the discrete logarithm problem in practice.