Improve the performance of roaring bitmap under sparse cases

In a sparse roaringbitmap, the size of array_container may often be 1, 2, 3. But we can actually reuse the container pointer (sizeof(void*)) as an array of uint16_t to avoid extra memory allocation.
eg:

typedef struct roaring_array_s {
    int32_t size;
    int32_t allocation_size;
    void **containers;
    uint16_t *keys;
    uint8_t *typecodes;
    uint8_t flags;
} roaring_array_t;

roaring_array_t ra;
// some init code
ra.typecodes[i] = SINGLE_CONTAIN_TYPE + cardinality << 6;
((uint16_t*)ra.containers[i])[1] = value1;
((uint16_t*)ra.containers[i])[2] = value2;
((uint16_t*)ra.containers[i])[3] = value3;

Start by tracing the roaring_array_t container and typecode handling described in the issue, then inspect how sparse array containers are allocated and accessed. Measure sparse cases before changing anything, and confirm that the implementation reduces allocation overhead while preserving container values and existing behavior.