llama: fold the random-access prefetch into its own feature flag

LLAMA_MMAP_RANDOM_PREFETCH existed to measure the two halves of the feature
apart, and the measurement is done: on a cold cache over the same wikitext
run, MADV_RANDOM without the batched readahead takes 94.4 s against 36.7 s
for an untouched mapping, while the pair together take 34.1 s. Suppressing
the kernel's readahead only pays if we replace it, so the split let a user
select a 2.6x regression through a documented switch.

Keep the accessor, since the call site reads better than a mode comparison,
but derive it from the mode alone.

src/llama-mmap.cpp

@@ -457,14 +457,7 @@ llama_mmap_random_mode llama_mmap_random_mode_get() {
 }
 
 bool llama_mmap_random_prefetch_enabled() {
-    // on with the feature, so the batched readahead that pays for the random hints cannot be
-    // left off by accident. separate only so the two halves can be measured apart.
-    static const bool enabled = []() {
-        const char * env = getenv("LLAMA_MMAP_RANDOM_PREFETCH");
-        return env == nullptr ? true : strcmp(env, "0") != 0;
-    }();
-
-    return llama_mmap_random_mode_get() != LLAMA_MMAP_RANDOM_OFF && enabled;
+    return llama_mmap_random_mode_get() != LLAMA_MMAP_RANDOM_OFF;
 }
 
 static size_t llama_mmap_page_size() {

src/llama-mmap.h

@@ -87,7 +87,9 @@ enum llama_mmap_random_mode {
 
 llama_mmap_random_mode llama_mmap_random_mode_get();
 
-// whether batched readahead ahead of a sparse gather is enabled (LLAMA_MMAP_RANDOM_PREFETCH)
+// batched readahead ahead of a sparse gather. not separately switchable: MADV_RANDOM suppresses
+// the kernel's own readahead, so without this the gather takes a synchronous fault per row and
+// runs 2.6x slower than leaving the mapping alone
 bool llama_mmap_random_prefetch_enabled();
 
 struct llama_mlock {