There's several forces behind this. Applications are more likely to be web-based or services - i.e. inherently distributed - and less likely to be desktop or traditional client/server, where almost all compute happened on a single server. As distributed services, maximizing statelessness and avoiding mutable shared memory is key to solving a lot of problems: scaling out (no shared memory), redundancy (keep a second copy of the application logic running somewhere, no sync required), recovery and idempotence (if something fails, try again until it succeeds - repeated attempts are safe).
Reliable persistent queues are part of that. They let you bring services up and down and switch over without down time, or restart after a failure and resume where they left off.
The problems of shared mutable state are best kept in specialized applications: databases, queuing systems, distributed caches, consistent key-value stores. Keeping state consistent in a distributed system is a genuinely hard problem, and STM isn't much help, except perhaps as an implementation detail in some of those specialized applications.
Web applications aren't special in this regard. We've just normalized being incredibly wasteful of computing resources when designing them, spinning up huge clusters to serve a workload that could be easily satisfied with a single machine (or a few machines for availability).
Premature scale-out, because we've forgotten how to scale-up, is the root of much evil.
Basically, poor STM performance in .Net is too much mutation and mutation-induced logging. This is why clojure's STM is much more widely used.
I have particularly deep experience in various Haskell projects and I must insist that no relatively (1) big Haskell project (2) with a hint of concurrency and (3) developed by several people can go without use of STM.
I can use MVars myself for problem space discovery. When there are at least two of use, I will use TVar and insist others should too.
I am not free to name the product, but I worked on a product that has done billions of dollars in revenue that designed a bunch of state management checkpointing using a custom STM implementation for C (yes, C). It made so, so many things straightforward in terms of corner cases; we also extended transactionality to sync over the wire in some cases.
I also think STM-adjacent designs are gaining some traction - for example, FoundationDB looks an awful lot like an STM system from the point of view of applications, much more than it looks like a traditional database.
1. It's become very common to structure software as a stateless service cluster backed by a stateful database technology. Most of the databases used in this approach would not be described as STM, but the net effect to the application is similar, even if involving more layers and complexity.
2. People have gotten pretty comfortable with simple mutex patterns connected by queues of various sorts. This is a sort of worse is better situation, where the simplicity and high performance of a mutex protected whatever in a toy test or microbenchmark is far more efficient than STM. However, a system composed of many of these mutex protected islands proves a whole different beast indeed. STM has largely been criticized from the perspective of the former, vs the latter.
There are many people who have made the observation that transactional concurrency control, state of the art garbage collection, and even file systems have been converging on similar features. This is being driven by the common constrains on both sides of what hardware and humans expect. In particular with persistent memory, I think you'll see all three of these unify into a single design, because systems that attempt to solve these problems separately will have very inferior match to the hardware.