Vec::extend/from_iter: accumulate in bulk instead of per-element push_back

[leighmcculloch]

What

The Extend impl for Vec (used by Vec::from_iter and .extend(..))
appended items by calling vec_push_back once per element. Each call is a host
call that also allocates a new intermediate vec object on the host. This change
accumulates items into a fixed-size stack buffer and flushes each full buffer
with a single bulk vec_new_from_slice + append.

Why

Same class of inefficiency that was fixed for BytesN::from in #1888
(per-element host calls where a bulk host op exists).

Net cost (baseline subtracted), measured on the guest budget with a WASM
compute-budget bench built the same way as the #1888 bench (baseline_* twin
subtracted; from_array shown as the already-bulk reference):

Bench	Before CPU	After CPU	Δ%	Before Mem	After Mem	Δ%
vec_from_iter_192	466,056	230,932	−50%	171,368	8,456	−95%

Bench contract and harness

Contract (one test_* cdylib crate):

#[no_std]
use soroban_sdk::{contract, contractimpl, Env, Vec};

#[contract]
pub struct Contract;

fn fp(v: &Vec) -> u32 {
    let l = v.len();
    if l == 0 { 0 } else { v.get_unchecked(0).wrapping_add(v.get_unchecked(l - 1)) }
}

#[contractimpl]
impl Contract {
    pub fn baseline_192(_e: Env) -> u32 { let d = [7u32; 192]; d[0].wrapping_add(d[191]) }
    pub fn vec_from_iter_192(e: Env) -> u32 { fp(&Vec::from_iter(&e, 0..192u32)) }
    pub fn vec_from_array_192(e: Env) -> u32 { fp(&Vec::from_array(&e, [7u32; 192])) }
}

Harness (in soroban-sdk/src/tests, #[ignore]d): register the wasm,
reset_unlimited(), call client.(), then print
cpu_instruction_cost() / memory_bytes_cost().

Run with make build-test-wasms then
cargo test --release -p soroban-sdk --lib --features testutils -- --ignored --nocapture.

Notes

The chunked approach works for any iterator and uses a fixed-size stack buffer
of Vals, so it works in no_std without alloc (the common contract case).
The from_iter doc note that claimed per-element host calls is updated. A unit
test (test_vec_from_iter_chunking) covers empty, sub-chunk,
exact-chunk-multiple and cross-chunk lengths.

Following the precedent of #1888, only the fix + correctness unit test are
committed here; the bench lives in this description rather than in the tree.

One of three related findings from sweeping the SDK for per-element host-call
patterns (see also: Vec::from_slice #1891, Vec::extend_from_slice #1892).