New README.md for llvm_version.txt (#938)

Fixes #936

build_tools/README.md

@@ -0,0 +1,34 @@
+## How to build LLVM for tpp-mlir
+
+The [instructions](https://github.com/plaidml/tpp-mlir?tab=readme-ov-file#how-to-build-llvm) to build LLVM are meant for cases where you only need to build once. 
+In those cases, a sub module would be more helpful, but those cases are rare, with so many other cases where sub modules would be detrimental.
+
+The main issues with using sub modules are:
+* Cloning the LLVM repo takes a long time, and it's redundant if you already have it elsewhere (very likely).
+* Building LLVM is an order of magnitude longer than tpp-mlir, and you probably already have the build elsewhere (personal or CI).
+* CI would have one LLVM copy (source + build) per builder (different directories).
+* Incompatible with [worktree](https://github.com/plaidml/tpp-mlir/wiki/Multiple-LLVM-Workspaces), which is the most sane way of developing in LLVM.
+
+The main way we use `llvm_version.txt` are listed below.
+
+### CI loop
+
+In CI we have a separate LLVM build clones the PR's version and installs in a shared disk, where any tpp-mlir build (CI or not) can use.
+This build has a _mutex_ so it only happens once per new hash. 
+The install dir is the hash of the last commit, which means CI can run concurrently on builds that are trying new versions as well as old ones.
+
+### Personal build in the cluster
+
+The build and compute nodes have access to the shared drive, so if you have access to the CI cluster, you don't need to build LLVM at all.
+Non-Intel folks can reproduce this on their local build server, using our CI scripts to build LLVM and install in a shared directory, to be used by multiple users.
+
+### Local builds
+
+If you are hacking on tpp-mlir, chances are you are also hacking on LLVM directly, and already have an LLVM tree (worktree even) elsewhere, and a build laying around.
+You can create a branch for tpp-mlir and install in a local directory (with the commit hash as name) and set-up your tpp-mlir build to pick up on that.
+This also helps when you go back and forth in tpp-mlir crossing LLVM version boundaries, there's no need to rebuild anything, as the previous hash install will still be there.
+
+### Benchmark builds
+
+If you're trying to reproduce our numbers from the paper, then you'll have a bit more work than recurring users to get everything built.
+Luckily, we have written a [script](https://github.com/plaidml/tpp-mlir/tree/main/scripts/benchmarks) for that task, which will do most of the manual steps for you, and get the benchmarks running.