r620: revamped thread-local memory management

* Don't preallocate sdust_buf or minizer list. kalloc should be fast enough -
  benchmarks needed to confirm.

* Fixed a memory leak caused by divergence estimate (post v2.5)

* Reset the kalloc buffer after mapping a long query. This reduces peak memory
  when large chunks of memory are allocated, at the cost of performance, though.

kalloc.h

@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
 typedef struct {
-	size_t capacity, available, n_blocks, n_cores;
+	size_t capacity, available, n_blocks, n_cores, largest;
 } km_stat_t;
 
 void *kmalloc(void *km, size_t size);