fix(gpu_prover): fix swapped tree cache mode in allocate_only_evaluation, disable setup recomputations (#173)

## What ❔

This PR fixes a bug that caused unnecessary allocations. This PR also
disables setup coset re-computations.

## Why ❔

The bug cause additional unnecessary allocations, because of memory
issues we enabled setup re-computations previously, with this fix we can
disable them again, getting a 5-10 % raw proving performance boost.

## Is this a breaking change?
- [ ] Yes
- [x] No

## Checklist

- [x] PR title corresponds to the body of PR (we generate changelog
entries from PRs).
- [x] Code has been formatted.

Author: robik75

gpu_prover/src/execution/gpu_worker.rs

@@ -128,7 +128,7 @@ fn gpu_worker(
                         circuit,
                         log_lde_factor,
                         log_tree_cap_size,
-                        true,
+                        false,
                         setup_trees_and_caps,
                         &context,
                     )?;

gpu_prover/src/prover/proof.rs

@@ -124,6 +124,8 @@ pub fn prove<'a, A: GoodAllocator>(
     setup_range.start(stream)?;
     setup.ensure_is_extended(context)?;
     setup_range.end(stream)?;
+    #[cfg(feature = "log_gpu_mem_usage")]
+    context.log_gpu_mem_usage("after setup.ensure_is_extended");
 
     let mut stage_1_output = StageOneOutput::allocate_trace_holders(
         &circuit,

gpu_prover/src/prover/trace_holder.rs

@@ -363,9 +363,9 @@ impl<T> TraceHolder<T> {
             false => CosetsHolder::Full(vec![evaluations]),
         };
         let trees = match trees_cache_mode {
-            TreesCacheMode::CacheNone => TreesHolder::Full(vec![]),
+            TreesCacheMode::CacheNone => TreesHolder::None,
             TreesCacheMode::CachePatrial => unimplemented!(),
-            TreesCacheMode::CacheFull => TreesHolder::None,
+            TreesCacheMode::CacheFull => TreesHolder::Full(vec![]),
         };
         Ok(Self {
             log_domain_size,