ja: use the Japanese mathematical terms

[yasumorishima]

Follows #728 on the ja branch. This one is terminology only — no rule structure changes.

The seeding reached for a loanword or a dictionary gloss where Japanese mathematics has a settled term. Some of the results are not words at all (アドフゲート for adjugate), and some are real words that mean something else (規範 is a social norm; 限界 is a bound or a ceiling; 上記 means "the aforementioned", not "above").

definitions.yaml

	was	now
limit	限界として (as a bound)	極限
tends to	傾向がある (has a tendency)	に近づく
tends to from above	上記から (from the aforementioned)	上から近づく
tends to from below	以下に傾向がある	下から近づく
set / empty set / set difference	セット / 空のセット / セットの違い	集合 / 空集合 / 差集合
subset / superset	サブセット / スーパーセット	部分集合 / 上位集合
mode	モード	最頻値
matrix / identity matrix	マトリクス / アイデンティティ マトリックス	行列 / 単位行列
adjugate	アドフゲート (not a word)	余因子行列
norm / end norm	規範 (social norm) / エンド規範	ノルム / ノルム閉じ
transpose	トランスポーズ	転置
by (dimensions)	によって (by means of)	かける

Two of these needed checking rather than guessing:

• adjugate → 余因子行列. The obvious rendering 随伴行列 is wrong: that is the Hermitian adjoint. Japanese 余因子行列 denotes the adjugate, even though a literal back-translation of 余因子 is "cofactor" — the naming convention differs between the two languages.
• augmented matrix → 拡大行列 rather than 拡大係数行列, since the latter presupposes the coefficient matrix of a linear system and MathCAT speaks the phrase on its own.

The same words elsewhere

definitions.yaml was not enough — the same terms appear in the rule files. Sweeping them found 20 more occurrences of the matrix loanwords and 11 of によって in SharedRules/general.yaml and overview.yaml, and the set words in ClearSpeak_Rules.yaml, SimpleSpeak_Rules.yaml and unicode.yaml. Also fixed there: ゼロマトリックス → 零行列, 対角マトリックス → 対角行列, 端のマトリックス → 行列終了, 拡張されたマトリックス → 拡大行列, and the number sets (複雑な数字 → 複素数, 自然な数字 → 自然数, 合理的な番号 → 有理数, インテガーズ → 整数).

navigate.yaml deliberately keeps its two セット. There "set" is the verb — "set the value of the next entry" — so 集合 would be wrong. A blanket replacement would have broken it.

Checks

• audit-translations ja: 28 rule differences and 3682 untranslated, unchanged from ja. Nothing structural moved.
• Test expectations are the actual CI output. Added coverage for 集合, ノルム, 最頻値, に近づく and the matrix reading.

Not in this PR

unicode.yaml still reads × as によって when ClearSpeak_MultSymbolX is By. Japanese does not separate "by" from "times" for the multiplication sign, so setting that to かける would make the By preference identical to Auto — that is a decision about the preference, not a translation fix, so I left it.

The intent readings come out as 集合 の x, 最頻値 の x and so on. That trailing の is the same connective #726 replaced with オブ in the large-operator rules, and it has the same problem here, but it lives in the function-application rule rather than in these definitions.